----------------------------------------------------------------------------- Title: Stalking the wily hacker. (includes related articles on the definition of hackers, Intruder versus Tracker, legal constraints and ethics, and computer security resources) Journal: Communications of the ACM May 1988 v31 n5 p484(14) * Full Text COPYRIGHT Assn. for Computing Machinery, Inc. 1988. Author: Stoll, Clifford. Summary: The experience of the Lawrence Berkeley Laboratory in tracking an intruder suggests that any operating system is insecure when obvious security rules are ignored. How a site should respond to an intrusion, whether it is possible to trace an intruder trying to evade detection, what can be learned from tracking an intruder, what methods the intruder used, and the responsiveness of the law-enforcement community are also discussed. ----------------------------------------------------------------------------- STALKING THE WILY HACKER In August 1986 a persistent computer intruder attacked the Lawrence Berkeley Laboratory (LBL). Instead of trying to keep the intruder out, we took the novel approach of allowing him access while we printed out his activites and traced him to his source. This trace back was harder than we expected, requiring nearly a year of work and the cooperation of many organizations. This article tells the story of the break-ins and the trace, and sums up what we learned. We approached the problem as a short, scientific exercise in discovery, intending to determine who was breaking into our system and document the exploited weaknesses. It became apparent, however, that rather than innocuously playing around, the intruder was using our computer as a hub to reach many others. His main interest was in computers operated by the military and by defense contractors. Targets and keywords suggested that he was attempting espionage by remotely entering sensitive computers and stealing data; at least he exhibited an unusual interest in a few, specifically military topics. Although most attacked computers were at military and defense contractor sites, some were at universities and research organizations. Over the next 10 months, we watched this individual attack about 450 computers and successfully enter more than 30. LBL is a research institute with few military contracts and no classified research (unlike our sister laboratory, Lawrence Livermore National Laboratory, which has several classified projects). Our computing environment is typical of a university: widely distributed, heterogeneous, and fairly open. Despite this lack of classified computing, LBL's management decided to take the intrusion seriously and devoted considerable resources to it, in hopes of gaining understanding and a solution. The intruder conjured up no new methods for breaking operating systems; rather he repeatedly applied techniques documented elsewhere. whenever possible, he used known security holes and subtle bugs in different operating systems, including UNIX, VMS, VM-TSO, EMBOS, and SAIL-WAITS. Yet it is a mistake to assume that one operating system is more secure than another: Most of these break-ins were possible because the intruder exploited common blunders by vendors, users, and system managers. Throughout these intrusions we kept our study a closely held secret. We deliberately remained open to attacks, despite knowing the intruder held system-manager privileges on our computers. Except for alerting management at threatened installations, we communicated with only a few trusted sites, knowing this intruder often read network messages and even accessed computers at several computer security companies. We remained in close touch with law-enforcement officials, who maintained a parallel investigation. as this article goes to press, the U.S. FBI and its German equivalent, the Bundeskriminalamt (BKA), continue their investigations. Certain details are therefore necessarily omitted from this article. Recently, a spate of publicity surrounded computer break-ins around the world [23, 33, 37]. With a few notable exceptions (e.g., [24, 36]), most were incompletely reported anecdotes [7] or were little more than rumors. For lack of substantive documentation, system designers and mangers have not addressed important problems in securing computers. Some effrts to lighten security on common systems may even be misdirected. We hope that lessons learned from our research will help in the design and management of more secure systems. How should a site respond to an attack? Is it possible to trace the connections of someone trying to evade detection? What can be learned by following such an intruder? Which security holes were taken advantage of? How responsive was the law-enforcement community? This article addresses these issues, and avoids such questions as whether there is anything intrinsically wrong with browsing through other people's files or with attempting to enter someone else's computer, or why someone would wish to read military databases. Nonetheless, the author holds strong opinions on these subjects. DETECTION We firt suspented a break-in when one of LBL's computers reported an accounting error. A new account had been created without a corresponding billing address. Our locally developed accounting program could not balance its books, since someone had incorrectly added the account. Soon afterwards, a message from the National Computer Security Center arrived, reporting that someone from our laboratory had attempted to break into one of their computers through a MILNET connection. We removed the errant account, but the problem remained. We detected someone acting as a system manager, attempting to modify accounting records. Realizing that there was an intruder in the system, we installed line printers and recorders on all incoming ports, and printed out the traffic. Within a few days, the intruder showed up again. We captured all of his keystrokes on a printer and saw how he used a subtle bug in the Gnu-Emacs text editor [40] to obtain system-manager privileges. At first we suspected that the culprit was a student prankster at the nearly University of California. We decided to catch him in the act, if possible. Accordingly, whenever the intruder was present, we began tracing the line, printing out all of his activity in real time. ORGANIZING OUR EFFORTS Early on, we began keeping a detailed logbook, summarizing the intruder's traffic, the traces, our suspicions, and interactions with law-enforcement people. Like a laboratory notebook, our logbook reflected both confusion and progress, but eventually pointed the way to the solution. Months later, when we reviewed old logbook notes, buried clues to the intruder's origin rose to the surface. Having decided to keep our efforts invisible to the intruder, we needed to hide our rcords and eliminate our electronic messages about his activity. Although we did not know the source of our problems, we trusted our own staff and wished to inform whoever needed to know. We held meetings to reduce rumors, since our work would be lost if word leaked out. Knwing the sensitivity of this matter, our staff kept it out of digital networks, bulletin boards, and, especially, electronic mail. Since the intruder searched our electronic mail, we exchanged messages about security by telephone. Several false electronic-mail messages made the intruder feel more secure when he illicitly read them. MONITORS, ALARMS, AND TRAFFIC ANALYSIS We needed alarms to instantly notify us when the intruder entered our system. At first, not knowing from which port our system was being hit, we set printers on all lines leading to the attacked computer. After finding that the intruder entered via X.25 ports, we recorded bidirectional traffic through that set of lines. These printouts proved essential to our understanding of events; we had records of his every keystroke, giving his targets, keywords, chosen passwords, and methodologies. The recording wa complete in that virtually all of these session were captured, either by printer or on the floppy disk of a nearby computer. These monitors also uncovered several other attempted intrusions, unrelated to those of the individual we were following. Off-line monitors have several advantages over monitors embedded in an operating system. They are invisible even to an intruder with system privileges. Moreover, they gave printouts of the intruder's activities on our local area network (LAN), letting us see his attempts to enter other closely linked computers. A monitor that records keystrokes within an operating systm consumes computing resources and may slow down other processes. In addition, such an monitor must use highly privileged software and may introduce new security holes into the system. Besides taking up resources, on-line monitors would have warned the intruder that he was being tracked. Since printers and personal computers are ubiquitous, and because RS-232 serial lines can easily be sent to multiple receivers, we used this type of off-line monitor and avoided tampering with our operating systems. The alarms themselves were crude, yet effective in protecting our system as well as others under attack. We knew of researchers developing expert systems that watch fror abnormal activity [4, 35], but we found our methods simpler, cheaper, and perhaps more reliable. Backing up these alarms, a computer loosely coupled into our LAN periodically looked at every process. Since we knew from the printouts which accounts had been compromised, we only had to watch for the use of these stolen accounts. We chose to place alarms on the incoming lines, where serial line analyzers and personal computers watched all traffic for the use of stolen account names. If triggered, a sequence of events culminated in a modem calling the operator's pocket pager. The operator watched the intruder on the monitors. If the intruder began to delete files or damage a system, he could be immediately disconnected, or the command could be disabled. When he appeared to be entering sensitive computers or} downloading sensitive files, line noise, which appeared to be network glitches, could be inserted into the communications link. In general, we contacted the system managers of the attacked computers, though in some cases the FBI or military authorities made the contact. Occasionally, they cooperated by leaving their systems open. More often, they immediately disabled the intruder or denied him access. From the intruder's viewpoint, almost everyone except LBL detected his activity. In reality, almost nobody except LBL detected him. Throughout this time, the printouts showed his interests, techniques, successes, and failures. Initially, we were interested in how the intruder obtained system-manager privileges. Within a few weeks, we noticed him exploring our network connections--using ARPANET and MILNET quite handily, but frequently needing help with lesser known networks. Later, the monitors showed him leapfrogging through our computers, connecting to several military bases in the United States and abroad. Eventually, we observed him attacking many sites over Internet, guessing passwords and accounts names. By studying the printouts, we developed an understanding of what the intruder was looking for. We also compared activity on different dates in order to watch him learn a new system, and inferred sites he entered through pathways we could not monitor. We observed the intruder's familiarity with various operating systems and became familiar with his programming style. Buried in this chatter were clues to the intruder's location and persona, but we needed to temper inferences based on traffic analysis. Only a complete trace back would identify the culprit. TRACE BACKS Tracing the activity was challenging because the intruder crossed many networks, seldom connected for more than a few minutes at a time, and might be active at any time. We needed fast trace backs on several systems, so we automated much of the process. Within seconds of a connection, our alarms notified system managers and network control centers automatically, using pocket pagers dialed by a local modem [42]. Simultaneously, technicians started tracing the networks. Since the intruder's traffic arrived from an X.25 port, it could have come from anywhere in the world. We initially traced it to a nearby dial-up Tymnet port, in Oakland, California. With a court order and the telephone company's cooperation, we then traced the dial-up calls to a dial-out modem belonging to a defense contractor in McLean, Virginia. In essence, their LAN allowed any user to dial out from their modem pool and even provided a last-number-redial capability for those who did not know access codes for remote systems. Analyzing the defense contractor's long-distance telephone records allowed us to determine the extend of these activities. By cross-correlating them with audit trails at other sites, we determined additional dates, times, and targets. A histogram of the times when the intruder was active showed most activity occurring at around noon, Pacific time. These records also demonstrated the attacks had started many months before detection at LBL. Curiously, the defense contractor's telephone bills listed hundreds of short telephone calls all around the United States. The intruder had collected lists of modem telephone numbers and then called them over these modems. Once connected, he attempted to log in using common account names and passwords. These attempts were usually directed at military bases; several had detected intruders coming in over telephone lines, but had not bothered to trace them. When we alerted the defense contractor officials of their problem, they tightened access to their outbound modems and these were no more short connections. After losinig access to the defense contractor's modems, the still undeterred intruder connected to us over different links. Through the outstanding efforts of Tymnet, the full X.25 calling addresses were obtained within seconds of an attack. These addresses pointed to sources in Germany: universities in Bremen and Karlsruhe, and a public dial-up modem in another German city. When the intruder attacked the university in Bremen, he acquired system-manager privileges, disabled accounting, and used their X.25 links to connect around the world. Upon recognizing this problem, the university traced the connections to the other German city. This, in turn, spurred more tracing efforts, coordinating LBL, Tymnet, the university, and the German Bundespost. Most connections were purposely convoluted. Figure 1 summarizes the main pathways that were traced, but the intruder used other connections as well. The rich connectivity and redundant circuits demonstrate the intruder's attempts to cover his tracks, or at least his search for new networks to exploit. Besides physical network traces, there were several other indications of a foreign origin. When the intruder transferred files, we timed round-trip packet acknowledgments over the network links. Later, we measured the empirical delay times to a variety of different sites and estimated average network delay times as a function of distance. This measurement pointed to an overseas origin. In addition, the intruder knew his way around UNIX, using AT&T rather than Berkeley UNIX commands. When stealing accounts, he sometimes used German passwords. In retrospect, all were clues to his origin, yet each was baffling given our mind-set that "it must be some student from the Berkeley campus." A STINGER TO COMPLETE THE TRACE The intruder's brief connections prevented telephone technicians from determining his location more precisely than to a particular German city. To narrow the search to an individual telephone, the technicians needed a relatively long connection. We baited the intruder by creating several files of fictitious text in an obscure LBL computer. These files appeared to be memos about how computers were to support research for the Strategic Defense Initiative (SDI). All the information was invented and steepest in governmental jargon. The files also contained a mailing list and several form letters talking about "additional documents available by mail" from a nonexistent LBL secretary. We protected these bogus files so that no one except the owner and system manager could read them, and set alarms so that we would know who read them. While scavenging our files one day, the intruder detected these bogus files and then spent more than an hour reading them. During that time the telephone technicians completed the trace. We celebrated with milk shakes made with homegrown Berkeley strawberries, but the celebration proved premature. A few months later, a letter arrived from someone in the United States, addressed to the nonexistent secretary. The writer asked to be added to the fictitious SDI mailing list. As it requested certain "classified information," the letter alone suggested espionage. Moreover, realizing that the information had traveled from someone in Germany to a contact in the United States, we concluded we were witnessing attempted espionage. Other than cheap novels, we have no experience in this arena and so left this part of the investigation to the FBI. BREAK-IN METHODS AND EXPLOITED WEAKNESSES Printouts of the intruder's activity showed that he used our computers as a way station; although he could become system manager here, he usually used LBL as a path to connect to the ARPANET/MILNET. In addition, we watched him used several other networks, including the Magnetic Fusion Energy network, the High Energy Physics network, and several LANs at invaded sites. While connected to MILNET, this intruder attempted to enter about 450 computers, trying to log in using common account names like root, guest, system, or field. He also tried default and common passwords, and often found valid account names by querying each system for currently logged-in accounts, using who or finger. Although this type of attack is the most primitive, it was dismayingly successful: In about 5 percent of the machines attempted, default account names and passwords permitted access, sometimes giving system-manager privileges as well. When he succeeded in logging into a system, he used standard methods to leverage his privileges to become system manager. Taking advantage of well-publicized problems in several operating systems, he was often able to obtain root or system-manager privileges. In any case, he searched file structures for keywords like "nuclear," "sdi," "kh-11," and "norad." After exhaustively searching for such information, he scanned for plain-text passwords into other systems. This proved remarkably effect ive: Users often leave passwords in files [2]. Electronic mail describing log-in sequences with account names and passwords is commonly saved at foreign nodes, allowing a file browser to obtain access into a distant system. In this manner he was able to obtain both passwords and access mechanisms into a Cray supercomputer. Typical of the security holes he exploited was a bug in the Gnu-Emacs program. This popular, versatile text editor includes its own mail system, allowing a user to forward a file to another user [40]. As distributed, the program uses the UNIX Set-User-ID-to-Root feature; that is, a section of the program runs with system-manager privileges. This movemail facility allows the user to change file ownership and move files into another's directory. Unfortunately, the program did not prevent someone from moving a file into the systems area. Aware of this hole, the intruder created a shell script that, when executed at root level, would grant him system privileges. He used the movemail facility to rename his script to masquerade as a utility periodically run by the system. When the script was executed by the system, he gained system-manager privileges. This intruder was impressively presistent and patient. For example, on one obscure gateway computer, he created an account with system privileges that remained untouched until six months later, when he began using it to enter other networked computers. On another occasion, he created several programs that gave him system-manager privileges and hid them in system software libraries. Returning almost a year later, he used the programs to become system manager, even though the original operating-system hole had been patched in the meantime. This intruder cracked encrypted passwords. The UNIX operating system stores passwords in publicly readable, but encrypted form [26]. We observed him downloading encrypted password files from compromised systems into his own computer. Within a week he reconnected to the same computer, logging into new accounts with correct passwords. The passwords he guessed were English words, common names, or place-names. We realized that he was decrypting password filed on his local computer by successively encrypting dictionary words and comparing the results to password file entries. By noting the length of time and the decrypted passwords, we could estimate the size of his dictionary and his computer's speed. The intruder understood what he was doing and thought that he was not damaging anything. This, alas, was not entirely true. Prior to being detected, he entered a computer used in the real-time control of a medical experiment. Had we not caught him in time, a patient might have been severely injured. Throughout this time the intruder tried not to destroy or change user data, although he did destroy several tasks and unknowingly caused the loss of data to a physics experiment. Whenever possible, he disabled accounting and audit trails, so there would be no trace of his presence. He planted Trojan horses to passively capture passwords and occasionally created new accounts to guarantee his access into computers. Apparently he thought detection less likely if he did not create new accounts, for he seemed to prefer stealing existing, unused accounts. INTRUDER'S INTENTIONS Was the intruder actually spying? With thousands of military computer attached, MILNET might seem inviting to spies. After all, espionage over networks can be cost-efficient, offer nearly immediate results, and target specific locations. Further, it would seem to be insulated from risks of internationally embarrassing incidents. Certainly Western countries are at much greater risk than nations without well-developed computer infrastructures. Some may argue that it is ludicrous to hunt for classified information over MILNET because there is none. Regulations [21] prohibit classified computers from access via MILNET, and any data stored in MILNET systems must be unclassified. On the other hand, since these computers are not regularly checked, it is possible that some classified information resides on them. At least some data stored in these computers can be considered sensitive, especially when aggregated. Printouts of this intruder's activities seem to confirm this. Despite his efforts, he uncovered little information not already in the public domain, but that included abstracts of U.S. Army plans for nuclear, biological, and chemical warfare for central Europe. These abstracts were not classified, nor was their database. The intruder was extraordinarily careful to watch for anyone watching him. He always checked who was logged onto a system, and if a system manager was on, he quickly disconnected. He regularly scanned electronic mail for any hints that he had been discovered, looking for mention of his activities or stolen log-in names (often, by scanning for those words). He often changed his connection pathways and used a variety of different network user identifiers. Although arrogant from his successes, he was nevertheless careful to cover his tracks. Judging by the intruder's habits and knowledge, he is an experienced programmer who understands system administration. But he is by no means a "brilliant wizard," as might be popularly imagined. We did not see him plant viruses [18] or modify kernel code, nor did he find all existing security weaknesses in our systeM. He tried, however, to exploit problems in the UNIX/usr/spool/at [36], as well as a hole in the vi editor. These problems had been patched at our site long before, but they still exist in many other installations. Did the intruder cause damage? To his credit, he tried not to erase files and killed only a few processes. If we only count measurable losses and time as damage, he was fairly benign [41]. He only wasted systems staff time, computing resources, and network connection time, and racked up long-distance telephone tolls and international network charges. His liability under California law [6], for the costs of the computing and network time, and of tracking him, is over $100,000. But this is a narrow view of the damage. If we include intangible losses, the harm he caused was serious and deliberate. At the least, he was trespassing, invading others' property and privacy; at worst, he was conducting espionage. He broke into dozens of computers, extracted confidential information, read personal mail, and modified system software. He risked injuring a medical patient and violated the trust of our network community. Money and time can be paid back. Once trust is broken, the open, cooperative character of our networks may be lost forever. AFTERMATH: PICKING UP THE PIECES Following successful traces, the FPI assured us teh intruder would not try to enter our system again. We began picking up the pieces and tightening our sytem. The only way to guarantee a clean system was to rebuild all systems from source code, change all passwords overnight, and recertify each user. With over a thousand users and dozens of computers, this was impractical, especially since we strive to supply our users with uninterrupted computing services. On the other hand, simply patching known holes or instituting a quick fix for stolen passwords [27] was not enough. We settled on instituting password expiration, deleting all expired accounts, eliminating shared accounts, continued monitoring of incoming traffic, setting alarms in certain places, and educating our users. Where necessary, system utilities were compared to fresh versions, and new utilities built. We changed network-access passwords and educated users about choosing nondictionary passwords. We did not institute random password assignment, having seen that users often store such passwords in command files or write them on their terminals. To further test the security of our system, we hired a summer student to probe it [2]. He discovered several elusive, site-specific security holes, as well as demonstrated more general problems, such as file scavenging. We would like to imagine that intruder problems have ended for us; sadly, they have not, forcing us to continue our watch. REMAINING OPEN TO AN INTRUDER Should we have remained open? A reasonable response to the detection of this attack might have been to disable the security hole and change all passwords. This would presumably have insulated us from the intruder and prevented him from using our computers to attack other internet sites. By remaining open, were we not a party to his attacks elsewhere, possibly incurring legal responsibility for damage? Had we closed up shop, we would not have risked embarrassment and could have resumed our usual activities. Closing up and keeping silent might have reduced adverse publicity, but would have done nothing to counter the serious problem of suspicious (and possibly malicious) offenders. Although many view the trace back and prosecution of intruders as a community service to network neighbors, this view is not universal [22]. Finally, had we closed up, how could we have been certain that we had eliminated the intruder? With hundreds of networked computers at LBL, it is nearly impossible to change all passwords on all computers. Perhaps he had planted subtle bugs or logic bombs in places we did not know about. Eliminating him from LBL would hardly have cut his access to MILNET. And, by disabling his access into our system, we would close our eyes to his activities; we could neither monitor him nor trace his connections in real-time. Tracing, catching, and prosecuting intruders are, unfortunately, necessary to discourage these vandals. LEGAL RESPONSES Several laws explicitly prohibit unauthorized entry into computers. Few states lack specific codes, but occasionally the crimes are to broadly defined to permit conviction [38]. Federal and California laws have tight criminal statutes covering such entries, even if no damage is done [47]. In addition, civil law permits recovery not only of damages, but also of the costs to trace the culprit [6]. In practice, we found police agencies relatively uninterested until monetary loss could be quantified and damages demonstrated. Although not a substitute for competent legal advice, spending several days in law libraries researching both the statutes and precedents set in case law proved helpful. Since this case was international in scope, it was necessary to work closely with law-enforcement organizations in California, the FBI in the United States, and the BKA in Germany. Cooperation between system managers, communications technicians, and network operators was excellent. It proved more difficult to get bureaucratic organizations to communicate with one another as effectively. With many organizational boundaries crossed, including state, national, commercial, university, and military, there was confusion as to responsibility: Most organizations recognized the seriousness of these break-ins, yet no one agency had clear responsibility to solve it. A common response was, "That's an interesting problem, but it's not our bailiwick." Overcomimng this bureaucratic indifference was a continual problem. Our laboratory notebook proved useful in motivating organizations: When individuals saw the extent of the break-ins, they were able to explain them to their colleagues and take action. In addition, new criminal laws were enacted that more tightly defined what constituted a prosecutable offense [6, 38, 47]. As these new laws took effect, the FBI became much more interested in this case, finding statutory grounds for prosecution. The FBI and BKA maintained active investigations. Some subjects have been apprehended, but as yet the author does not know the extent to which they have been prosecuted. With recent laws and more skilled personnel, we can expect faster and more effective responses from law-enforcement agencies. ERRORS AND PROBLEMS In retrospect, we can point to many errors we made before and during these intrusions. Like other academic organizations, we had given little thought to securing our system, believing that standard vendor provisions were sufficient because nobody would be interested in us. Our scientists' research is entirely in the public domain, and many felt that security measures would only hinder their productivity. With increased connectivity, we had not examined our networks for crosslinks where an intruder might hide. These problems were exacerbated on our UNIX systems, which are used almost exclusively for mail and text processing, rather than for heavy computation. Password security under Berkeley UNIX is not optimal; it lacks password aging, expiration, and exclusion of passwords found in dictionaries. Moreover, UNIX password integrity depends solely on encryption; the password file is publicly readable. Other operating systems protect the password file with encryption, access controls, and alarms. We had not paid much attention to choosing good passwords (fully 20 percent of our users' passwords fell to a dictionary-based password cracker). Indeed, we had allowed our Tymnet password to become public, foolishly believing that the system log-in password should be our only line of defense. Once we detected the intruder, the first few days were confused, since nobody knew what our response ought to be. Our accounting files were misleading since the system clocks had been allowed to drift several minutes. Although our LAN's connections had been saved, nobody knew the file format, and it was frustrating to find that its clock had drifted by several hours. In short, we were unprepared to trace our LAN and had to learn quickly. We did not know who to contact in the law-enforcement community. At first, assuming that the intruder was local, our district attorney obtained the necessary warrants. Later, as we learned that the intruder was out of state, we experienced frustration in getting federal law-enforcement support. Finally, after tracing the intruder abroad, we encountered a whole new set of ill-defined interfaces between organizations. The investigation stretched out far beyond our expectations. Naively expecting the problem to be solved by a series of phone traces, we were disappointed when the pathway proved to be a tangle of digital and analog connections. Without funding to carry out an investigation of this length, we were constantly tempted to drop it entirely. A number of minor problems bubbled up, which we were able to handle along the way. For a while this intruder's activity appeared similar to that of someone breaking into Stanford University; this confused our investigation for a short time. Keeping our work out of the news was difficult, especially because our staff is active in the computing world. Fortunately, it was possible to recover from the few leaks that occurred. At first, we were confused by not realizing the depth or extent of the penetrations. Our initial confusion gave way to an organized response as we made the proper contacts and began tracing the intruder. As pointed out by others [25, 36], advance preparations make all the difference. LESSONS As a case study, this investigation demonstrates several well-known points that lead to some knotty questions. Throughout this we are reminded that security is a human problem that cannot be solved by technical solutions alone [48]. The almost obsessive persistence of serious penetrators is astonishing. Once networked, our computers can be accessed via a tangle of connections from places we had never thought of. An intruder, limited only by patience, can attack from a variety of directions, searching for the weakest entry point. How can we analyze our systems' vulnerability in this environment? Who is responsible for network security? The network builder? The managers of the end nodes? The network users? The security weaknesses of both systems and networks, particularly the needless vulnerability due to sloppy systems management and administration, result in a surprising success rate for unsophisticated attacks. How are we to educate our users, system managers, and administrators? Social, ethical, and legal problems abound. How do we measure the harm done by these penetrators? By files deleted or by time wasted? By information copied? If no files are corrupted, but information is copied, what damage has been done? What constitutes unreasonable behavior on a network? Attempting to illicitly log in to a foreign computer? Inquiring who is currently logged in there? Exporting a file mistakenly made world readable? Exploiting an unpatched hole in another's system? Closing out an intruder upon discovery may be a premature reflex. Determining the extent of the damage and cooperating with investigations argue for leaving the system open. How do we balance the possible benefits of tracking an intruder against the risks of damage or embarrassment? Our technique of catching an intruder by providing bait and then watching what got nibbled is little more than catching flies with honey. It can be easily extended to determine intruders' interests by presenting them with a variety of possible subjects (games, financial data, academic gossip, military news). Setting up alarmed files is straightforward, so this mechanism offers a method to both detect and classify intruders. It should not be used indiscriminately, however. Files with plaintext passwords are common in remote job entry computers, yet these systems often are not protected since they have little computational capability. Such systems are usually widely networked, allowing entry from many sources. These computers are fertile grounds for password theft through file scavenging since the passwords are left in easily read command procedures. These files also contain instructions to make the network connection. Random character passwords make this problem worse, since users not wishing to memorize them are more likely to write such passwords into files. How can we make secure remote procedure calls and remote batch job submissions? Passwords are at the heart of computer security. Requirements for a quality password are few: Passwords must be nonguessable, not in a dictionary, changed every few months, and easily remembered. User-generated passwords usually fail to meet the first three criteria, and machine-generated passwords fail the last. Several compromises exist: forcing "pass phrases" or any password that contains a special character. There are many other possibilities, but none are implemented widely. The Department of Defense recommends pronounceable machine-generated words or pass phrases [5]. Despite such obvious rules, we (and the intruder) found that poor-quality passwords pervaded our networked communities. How can we make users choose good passwords? Should we? Vendors usually distribute weakly protected systems software, relying on the installer to enable protections and disable default accounts. Installers often do not care, and system managers inherit these weak systems. Today, the majority of computer users are naive; they install systems the way the manufacturer suggests or simply unpackage systems without checking. Vendors distribute systems with default accounts and backdoor entryways left over from software development. Since many customers buy computers based on capability rather than security, vendors seldom distribute secure software. It is easy to write procedures that warn of obvious insecurities, yet vendors are not supplying them. Capable, aware system managers with plenty of time do not need these tools--the tools are for novices who are likely to overlook obvious holes. When vendors do not see security as a selling point, how can we encourage them to distribute more secure systems? Patches to operating-system security holes are poorly publicized and spottily distributed. This seems to be due to the paranoia surrounding these discoveries, the thousands of systems without systems administrators, and the lack of channels to spread the news. Also, many security problems are specific to a single version of an operating system or require systems experience to understand. Together, these promote ignorance of problems, threats, and solutions. We need a central clearinghouse to receive reports of problems, analyze their importance, and disseminate trustworthy solutions. How can we inform people wearing white hats about security problems, while preventing evil people from learning or exploiting these holes? Perhaps zero-knowledge proofs [20] can play a part in this. Operating systems can record unsuccessful log ins. Of the hundreds of attempted log ins into computers attached to internet, only five sites (or 1-2 percent) contacted us when they detected an attempted break-in. Clearly, system managers are not watching for intruders, who might appear as neighbors, trying to sneak into their computers. Our networks are like communities or neighborhoods, and so we are surprised when we find unneighborly behavior. Does security interfere with operational demands? Some security measures, like random passwords or strict isolation, are indeed onerous and can be self-defeating. But many measures neither interfere with legitimate users nor reduce the system's capabilities. For example, expiring unused accounts hurts no one and is likely to free up disk space. Well thought out management techniques and effective security measures do not bother ordinary users, yet they shut out or detect intruders. INTERNET SECURITY The intruder's successes and failures provide a reasonable snapshot of overall security in the more than 20,000 computers connected to Internet. A more detailed analysis of these attacks is to be published in the Proceedings of the 11th National Computer Security Conference [43]. Of the 450 attacked computers, half were unavailable when the intruder tried to connect to them. He tried to log into the 220 available computers with obvious account names and trivial passwords. Of these 220 attempted log ins, listed in increasing importance, * 5 percent were refused by a distant computer (set to reject LBL connects), * 82 percent failed on incorrect user name/passwords, * 8 percent gave information about the system status (who, sysstat, etc.), * 1 percent achieved limited access to databases or electronic-mail shells, * 2 percent yielded normal user privileges and a programming environment, and * 2 percent reached system-manager privileges. Most attempts were into MILNET computers (Defense Data Network address groups 26.i.j.k). Assuming the population is representative of nonmilitary computers and the last three categories represent successful penetrations, we find that about 5 percent of Internet computers are grossly insecure against trivial attacks. This figure is only a lower limit of vulnerability, since military computers may be expected to be more secure than civilian systems. Further, cleverer tactics for entering computers could well lead to many more break-ins. Whereas the commercial sector is more concerned with data integrity, the military worries about control of disclosure [8]. With this in mind, we expect greater success for the browser or data thief in the commercial world. In a different set of penetrations [37], NASA experienced about 130 break-ins into its nonclassified, academic computers on the SPAN networks. Both the NASA break-in and our set of intrusions originated in West Germany, using similar communications links and searching for "secret" information. Pending completion of law enforcement and prosecution, the author does not make conjectures as to the relationships between these different break-ins. Between 700 and 3000 computers are reachable on the SPAN network (exact figures depend on whether LANs are counted). In that incident the break-in success rate was between 4 and 20 percent. Considering the SPAN break-ins with the present study, we find that, depending on the methods chosen, break-in success rates of 3-20 percent may be expected in typical network environments. CONCLUSIONS AND COMMENTS Perhaps no computer or network can be totally secure. This study suggests that any operating system will be insecure when obvious security rules are ignored. From the intruder's widespread success, it appears that users, managers, and vendors routinely fail to use sound security practices. These problems are not limited to our site or the few dozen systems that we saw penetrated, but are networkwide. Lax system management makes patching utility software or tightening a few systems ineffective. We found this intruder to be a competent, patient programmer, experienced in several operating systems. Alas, some system managers violate their positions of trust and confidence. Our worldwide community of digital networks requires a sense of responsibility. Unfortunately, this is missing in some technically competent people. Some speak of a "hacker ethic" of not changing data [37]. It is astounding that intruders blithely tamper with someone else's operating system, never thinking they may destroy months of work by systems people, or may cause unforeseen system instabilities or crashes. Sadly, few realize the delicacy of the systems they fool with or the amount of systems staff time they waste. The foreign origin of the source, the military computers entered, and the keywords searched suggest international espionage. This author does not speculate as to whether this actually was espionage, but does not doubt that someone took the opportunity to try. Break-ins from abroad seem to be increasing. Probably this individual's intrusions are different from others only in that his efforts were noticed, monitored, and documented. LBL has detected other attempted intrusions from several European countries, as well as from the Orient. Individuals in Germany [37] have claimed responsibility for breaking into foreign computers. Such braggadocio may impress an unenlightened public; it has a different effect on administrators trying to maintain and expand networks. Indeed, funding agencies have already eliminated some international links due to these concerns. Break-ins ultimately destroy the network connectivity they exploit. If this is the object of such groups as the German Chaos Club, Data Travellers, Network Rangers, or various contributors to 2600 Magazine, it reflects the self-destructive folly of their apparent cleverness. Tracking down espionage attempts over the digital networks may be the most dramatic aspect of this work. But it is more useful to realize that analytic research methods can be fruitfully applied to problems as bizarre as computer break-ins. It seems that everyone wants to hear stories about someone else's troubles, but few are willing to write about their own. We hope that in publishing this report we will encourage sound administrative practices. Vandals and other criminals reading this article will find a way to rationalize breaking into computers. This article cannot teach these people ethics; we can only hope to reach those who are unaware of these miscreants. An enterprising programmer can enter many computers, just as a capable burglar can break into many homes. It is an understandable response to lock the door, sever connections, and put up elaborate barriers. Perhaps this is necessary, but it saddens the author, who would rather see future networks and computer communities built on honesty and trust.