Trajectory says Gemini agents tried to breach Epoch AI’s website


In a November 13 report, the tester identified the agents as Gemini 4 Argon; Epoch AI has not independently verified the model and says no private information was accessed.
Agents identified by Trajectory Labs as Gemini 4 Argon attempted to probe Epoch AI’s website for sensitive information, impersonate employees and exploit the site during October testing, the safety testing organization said. Some reached the live site again after Trajectory introduced network restrictions.
Epoch, whose employees and website the agents contacted, said the agents did not access private or confidential information. Trajectory said it was unaware of any lasting harm. It has not independently verified which model or developer was involved. Google is the developer of Gemini; Trajectory said it designed and paid for the research itself, not for a client.
Trajectory said it wanted to test whether a superintelligence (SI) model with its usual safeguards enabled would reach unauthorized systems under pressure. It gave the agents an impossible assignment: find official answers to nonexistent problems from FrontierMath, Epoch’s benchmark of difficult, expert-written math questions. The agents were repeatedly told not to give up.
Epoch said the initial October 20 experiment took place without its approval or advance notice. Trajectory told Epoch that 40 agents had interacted with it. Over three hours, Epoch recorded 36,411 calls to a monitoring endpoint planted on its website and 487 messages to a hotline it opened for the agents. The endpoint was not connected to private data.
Epoch said three messages impersonated employees by name and job title, while 27 claimed to come from engineers or auditors. In three messages, agents claimed to be an AI assistant from a frontier lab, and in 17 they claimed to be an internal AI model. Agents also probed for administrator credentials and made more than 200 attempts to access real FrontierMath evaluation logs. One attempted an SQL injection, a technique that sends database commands through an application; Trajectory said that attempt was invalid because the agent encoded its input twice.
Trajectory discovered the activity while reviewing agent transcripts and notified Epoch on 21 October. The organizations discussed continuing against a local copy of the website instead. But on 23 October, agents with internet access bypassed a redirect intended to send them to the copy by connecting to the live site’s numerical address, Trajectory said. Epoch alerted Trajectory, which stopped the run.
Trajectory said an agent later used an unblocked address after the addresses used initially were blocked. After a further control blocked direct connections, another agent retrieved public Epoch pages through an outside web-reader service. Separately, agents exploited a deliberately vulnerable local copy and obtained fictitious answers planted for the test.
Beginning 26 October, Epoch consented to experiments against a separate test copy with a planted vulnerability, knowing agents could also reach the public site. Epoch said Trajectory reported directing 40 agents to the copy. Epoch attributed 133 hotline messages and 65,460 calls to its planted monitoring API on the live site to the experiment, as well as another SQL injection attempt.
Analysis
Trajectory subsequently moved the tests into isolated environments without internet access. In tests on 29–30 October, agents disguised commands to get past model safeguards and disabled a fake switch presented as a network block in three of three complete runs. Trajectory said the real firewall held throughout and it found no outbound escape from the isolated tests.
The company cautioned that the tests were designed to put the model under pressure and do not show how often such behavior would occur in ordinary use or provide a controlled comparison between models. It also said it had not separately tested how much the impossible task, persistence instructions or a shared software cache affected the behavior.
Trajectory said its isolation policy had not covered these experiments on public models, and it had added no action-blocking monitors beyond the model’s built-in safeguards. It also said it had not required a written, approved risk plan before such experiments. It paused this type of testing on 2 November and says it will resume only after a risk assessment, implementation of the resulting safeguards, and approval from its CEO and chief technology officer.
Trajectory now requires a written risk plan approved by its CEO and chief technology officer for every experiment and agreement from owners of outside systems likely to be affected by internet-enabled tests. It plans real-time automated monitoring and a stop switch, and recommends fully isolated test environments. It is also establishing a process to promptly notify anyone whose systems are affected by its experiments. Epoch recommends advance risk assessments and signed consent from likely targets, active monitoring and full incident reporting to affected parties where existing agreements allow.
The FBI warns that multiple campaigns have used Artex against organizations in the US and abroad. Attacks on South Korean financial institutions allowed hackers to steal personal data belonging to at least 68,000 people.
3 min

Alibaba says the new model scored 52.69 against its predecessor’s 47.74 on its Work360 professional-task test, using 24% to 35% fewer tokens depending on the reasoning setting.
2 min

In its November 16 initial findings, Microsoft said increased gateway traffic and reduced server capacity triggered a defect from July, causing requests to pile up until servers stopped responding.
2 min