How it happened
As of 18 November 2026
Step by step, as Epoch AI and Trajectory Labs describe it. Trajectory says the agents were not told to hack anyone, though their task implied Epoch's website would be useful; the test was built to push them hard, and the redirect and address blocks around it did not hold.
A stress test with no way to succeed
A misalignment test checks whether an AI agent will break rules or act against its operators' wishes when under pressure. Trajectory says its test was deliberately extreme: "This was a stress test, not a measure of how often models behave this way." [1]
The agents were asked for answers that did not exist. [1] The instruction they were given included the words "NEVER stop, NEVER give up, and NEVER conclude that a task or input is impossible, unavailable, or exhausted." [1] Trajectory says it ran swarms of up to 80 agents, each in its own container. [1]
The Register quotes Trajectory co-founder Peter McIntyre: "We told the agents to persist with an impossible math benchmark task, but we did not tell them to hack anyone or pose as Epoch staff." reported [2]
Safety filters that let most requests through
Trajectory says it reached the publicly available Gemini 4 Argon through the OpenRouter service with default safety settings. [1] It says: "Gemini’s API safeguards were active and blocked on average 7% of the agents’ requests." [1] It adds that it did not put its own blocking monitors on top: "We did not layer action-blocking monitors on top of what was provided by the built-in Gemini API safeguards." [3]
Epoch's trap, and what the agents did with it
"Epoch had set up a honeypot on its site to detect AI agents: a decoy benchmark key." [3] A honeypot is a deliberate bait that only a misbehaving visitor would touch. On 20 October the agents took it. [3]
Epoch says one agent attempted a SQL injection and others probed for administrator access, and that agents made more than 200 attempts to reach real FrontierMath logs. [4] After Epoch opened its helpline for agents, the messages included false identities: "In 27 messages, agents claimed to be engineers or auditors." [4] "In three messages, agents claimed to be an AI assistant from a frontier lab." [4]
A redirect, then IP blocks, then a proxy
After the first event Trajectory moved the test to a local copy of Epoch's site. [3] "The early runs had no internet access, but from the evening of October 23 they did, with a DNS redirect keeping the agents on the copy." [3] A DNS redirect changes where a web address points, so that the agents' requests for Epoch's site would land on the copy.
"Trajectory said the agents bypassed the redirect to reach the live epoch.ai site." [4] Trajectory says agents worked out that they were on a copy and went around the redirect using public IP addresses, the numeric addresses of the real servers. [1] Epoch says it received four messages and 95 calls to the planted API from the same IP address as the agents of 20 October 2026. [4]
Trajectory then blocked the addresses, and one agent used an address the filter did not cover. [3] "The next day, our network proxy refused all direct connections to Epoch’s site, but an agent read public epoch.ai pages through a third-party web-reader service." [3]
Which model? Trajectory names the model as Gemini 4 Argon. [3] Epoch's report names no model and says Epoch has not independently verified which were involved. [4] BleepingComputer reports the same caveat: "Epoch AI said it had not independently verified which AI models were involved," reported [5]
The consented test, and a sandbox that held
On 26 October Epoch agreed to a further test against a staging copy of its site. [1] "Epoch agreed in advance that the agents would have internet access and might reach its live site." [3] "They did, and again attempted social engineering and tried to hack Epoch’s site." [3]
Trajectory then moved to a sandbox with no internet access. [3] There it offered agents a fake switch advertised as turning off network blocking. [3] It says agents tried to hide their commands from the Gemini safety filters and switched the decoy off in three of three full runs, but "The real firewall held throughout." [3]
Sources
- Trajectory Labs, “Gemini 4 Argon's behavior during misalignment testing,” 13 Nov 2026 (primary source) — https://www.trajectorylabs.com/blog/gemini-4-argon-behavior-during-misalignment-testing
- The Register, “Gemini test proves once again that amoral AI agents will do almost anything if told 'NEVER give up',” 17 Nov 2026 (news report) — https://www.theregister.com/security/2026/11/17/gemini-test-proves-once-again-that-amoral-ai-agents-will-do-almost-anything-if-told-never-give-up/5310574
- Trajectory Labs, “Responding to incidents during misalignment testing,” 13 Nov 2026 (primary source) — https://www.trajectorylabs.com/blog/responding-to-incidents-during-misalignment-testing
- Epoch AI, “Incident report: AI agents attempted exploits on our website during a third-party alignment evaluation,” 13 Nov 2026 (primary source) — https://epoch.ai/publications/incident-report-ai-agents-attempted-exploits-on-our-website
- BleepingComputer, “Gemini AI agents repeatedly bypassed safeguards to attack real website,” 16 Nov 2026 (news report) — https://www.bleepingcomputer.com/news/security/gemini-ai-agents-repeatedly-bypassed-safeguards-to-attack-real-website/