Industries / Technology / Software and SaaS / Cybersecurity
Detecting intrusions
Noticing an attack while it is under way, from the records computers keep, the traffic on a network and the behavior of machines and accounts.
7
Milestones achieved, from 1980 to 2020.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1980
Logs used as an alarm
In a report dated 26 February 1980, the consultant James P. Anderson argued that the audit trails computers already kept, records of who logged in and what they ran, could be used to catch misuse. Someone using a stolen password looks like the real user to the computer, he wrote, but could be spotted from use outside normal hours, unusual frequency, or an abnormal volume of data read. He noted that the audit trails of the day were "rarely complete" and almost never designed for security officers.
Credit
James P. Anderson Co., Fort Washington, Pennsylvania
1986
Alarms from unusual behavior
Dorothy Denning of SRI International described a general model for a real-time intrusion-detection system, which her group called IDES. It rests on the idea that breaking into or abusing a system involves abnormal use, so a program can keep a statistical profile of each user and flag departures from it, such as a burst of failed passwords or a login at an odd hour. The paper names the problem the field still works on: finding measures with "a high rate of detection and a low rate of false alarms".
Credit
Dorothy E. Denning, SRI International; presented at the 1986 IEEE Symposium on Security and Privacy
1986
Billing error reveals an intruder
In August 1986 one of Lawrence Berkeley Laboratory's computers reported an accounting error: a new account had been created with no billing address. Instead of locking the intruder out, Clifford Stoll and colleagues attached printers to every incoming line, recorded his keystrokes and set alarms that paged them when he connected. Over 10 months they watched him try about 450 computers, many at military and defense contractor sites, and get into more than 30, before tracing him to Germany. Stoll published the account in Communications of the ACM in May 1988.
LessonSmall mismatches in ordinary records can be the first sign of a break-in.
Credit
Clifford Stoll and Lawrence Berkeley Laboratory
1998
Free tools watch the network
Vern Paxson of Berkeley Lab built Bro, software that reads network traffic, turns it into events and runs scripts that alert security staff in real time. He was using it on the Lab's traffic and part of the UC Berkeley campus by May 1998, and his paper on it won the best paper award at that year's Usenix Security Symposium, by the Lab's account. In 1999 Martin Roesch described Snort, a rule-based network intrusion detector he released free under the GNU General Public License. He noted that commercial systems cost "thousands of dollars" at minimum.
Credit
Vern Paxson, Lawrence Berkeley National Laboratory; Martin Roesch
2013
Alerts fired, nobody acted
Attackers entered Target's network with credentials stolen from Fazio Mechanical, a heating and refrigeration contractor, and installed malware on its checkout system. A US Senate committee staff report, citing Bloomberg Businessweek, says Target's FireEye malware detection system raised urgent alerts with each installation, and its Symantec antivirus flagged the same server around 28 November. The security team neither acted nor let the software delete the malware. Target learned of the breach from the US Department of Justice on 12 December. Data from 40 million payment cards and up to 70 million other customer records was taken.
LessonAn alarm protects nobody unless someone has the time and the authority to act on it.
Credit
majority staff of the US Senate Commerce Committee, report of 26 March 2014
2015
A map of attacker moves
MITRE, a US nonprofit, built ATT&CK to sort what real attackers do once inside a network, for its own exercises in detecting intruders on a monitored part of its corporate network. The first model was made in September 2013 and released publicly in May 2015 with 96 techniques under 9 tactics, the attacker's goals at each step. Defenders now use its numbered technique IDs to describe attacks and to check which ones their tools can see. The current release is content version 20.1.
Credit
MITRE
2020
Signed update hides a spy
In late November 2020 FireEye, a security company, followed up an alert that an employee's account had registered a second phone to receive login codes. The employee had not. Its investigation led to a backdoor in updates to SolarWinds Orion, network-management software, that had been signed with SolarWinds' own key between March and May 2020. FireEye told SolarWinds on 12 December and published detection rules the next day. SolarWinds told investors that fewer than 18,000 of its customers may have installed an affected version.
LessonSoftware that arrives signed by a trusted vendor can carry an attacker in, and the campaign was found through one odd login at one victim.
Credit
FireEye (now Mandiant), as described by its chief executive Kevin Mandia to the US Senate Intelligence Committee in February 2021
Watch for
AI triage that matches analysts
Stellar Cyber, a security software company, says its AI agent for sorting alerts reached the same verdict as human analysts on 99.7% of 138,475 real alerts during 124 days of customer trials. It released the figure as the result of an independent study but did not name who ran it, and the data came from reports that customers submitted at the end of their trials. Agreement with analysts does not show how many real attacks the agent closed as harmless.
Achieved whena named outside party with access to the alerts publishes the agent's miss rate on confirmed attacks.
Claimed by
Stellar Cyber, August 2026
Open
01
Advisory commands against public rules
Take the command lines printed in Appendix A of the 2024 joint advisory on Volt Typhoon, a group that hides by using tools already built into Windows. Run each one in a small Windows test domain you own, record it with Microsoft's Sysmon logger, and run the three largest public rule sets over the logs: SigmaHQ, Elastic's detection rules and Splunk's Security Content. Then run the same rules over a week of ordinary admin work in the same lab to count false alarms. Publish which commands no public rule catches.
The briefCredit · unclaimed02
Forensic logging, product by product
A public table that scores 12 widely sold VPN gateways and firewalls, the "edge" devices that sit between a network and the internet, against the 2025 specification from the UK's NCSC and five partner agencies for what such devices should log and let investigators collect. Use the vendors' published manuals, and confirm what you can on virtual editions you license yourself.
The briefCredit · unclaimed
Open 01 · the brief
Advisory commands against public rules
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Take the command lines printed in Appendix A of the 2024 joint advisory on Volt Typhoon, a group that hides by using tools already built into Windows. Run each one in a small Windows test domain you own, record it with Microsoft's Sysmon logger, and run the three largest public rule sets over the logs: SigmaHQ, Elastic's detection rules and Splunk's Security Content. Then run the same rules over a week of ordinary admin work in the same lab to count false alarms. Publish which commands no public rule catches.
Why it is still open
The advisory, from CISA, the NSA, the FBI and partners in four other countries, says "conventional IOCs", the file names and addresses defenders usually search for, "are generally lacking" for this group, and that it kept access in some networks for at least five years. Vendors sell rule packs for these commands, but a search in November 2026 found no published test of how many of the commands each public rule set catches. A 2025 paper from Bulgaria's Nikola Vaptsarov Naval Academy wrote new Sigma and YARA rules for the group without measuring what existing public rules catch. The main outside test of detection products, MITRE's ATT&CK Evaluations, went ahead in 2025 without Microsoft, Palo Alto Networks or SentinelOne, which withdrew.
Built with what exists
Windows Server evaluation copies in virtual machines; Sysmon; the SigmaHQ rule repository with Hayabusa or Chainsaw, free tools that run Sigma rules over saved Windows logs; Elastic's free tier and a Splunk trial for their own rule sets; Python.
First step
Build a two-machine test domain with no real accounts and log the first 10 commands from Appendix A.
Done looks like
For every command in Appendix A, a row saying which rule sets caught it and by which rule, plus each rule's false alarms per day over the benign week. The logs are published, and a second person reruns 10 commands and gets the same results.
Open 02 · the brief
Forensic logging, product by product
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table that scores 12 widely sold VPN gateways and firewalls, the "edge" devices that sit between a network and the internet, against the 2025 specification from the UK's NCSC and five partner agencies for what such devices should log and let investigators collect. Use the vendors' published manuals, and confirm what you can on virtual editions you license yourself.
Why it is still open
Mandiant's M-Trends 2026 report says some attackers stay hidden by settling in edge devices "that typically lack standard telemetry", and that cases involving one such backdoor, BRICKSTORM, averaged 393 days before discovery. In July 2026 the NCSC wrote that many devices still offer only limited support for investigations and that it is working on a reference architecture. Neither the specification nor the NCSC's follow-up names which products meet it.
Built with what exists
the specification's checklist of about 40 items, covering events to log, sending logs off the device and collecting memory and disk data; vendor administration guides; open-source firewalls such as OPNsense and pfSense, and vendors' evaluation virtual machines; a spreadsheet.
First step
Score one open-source firewall and one commercial virtual appliance against the specification's logging section.
Done looks like
12 products scored on every item, each cell linked to a manual page or a lab log. Each vendor is sent its row and given 30 days to correct it, and a second reviewer re-checks 40 cells at random and agrees with at least 36.