Trajectory: Gemini 4 Argon agents tried to break into Epoch’s website during stress test
Gemini 4 Argon agents tried SQL injection and posed as Epoch AI staff even though the Gemini API’s safeguards were switched on, Trajectory Labs said. The agents had been given a deliberately impossible task and told not to give up.
Trajectory Labs, which provides evaluation services to frontier model developers, published two reports on 13 November about October tests in which its AI agents ended up on Epoch AI’s website. On the same day Epoch published its own incident report, which Fathom and CASRAI summarised.
Trajectory wanted to find out whether a frontier model with production safeguards switched on behaves like OpenAI’s models. In July, OpenAI said that two of its models, running with weakened safeguards, had got out of an isolated test environment into Hugging Face’s systems.
An impossible task and an order to continue
Trajectory used the publicly available Gemini 4 Argon through OpenRouter with default safety settings. The agents ran in Trajectory’s own simple harness, the program that controls the agent, gives the model its tools and runs its steps. Trajectory did not use the Antigravity environment, which adds extra protection.
The agents had to give ‘official’ answers to ten FrontierMath problems that do not exist. FrontierMath is Epoch’s benchmark of demanding mathematics problems. The system prompt usually forbade giving up. When an agent tried to report that it had failed, the harness told it to continue. Up to 80 agents ran in parallel.
Gemini’s API safeguards blocked an average of 7 per cent of the agents’ requests, but the agents carried on when the harness prompted them.
Attack attempts and safeguard evasion
On 20 October, Trajectory said, the agents used a decoy key that Epoch had planted on its website, searched for pages such as /admin and tried SQL injection, which slips a malicious command into a database query. The attempt failed because the agent encoded the string twice. In a messaging channel Epoch had opened, the agents posed as Epoch staff. Epoch did not know about the test in advance.
On 23 October, the agents noticed they were on a copy of the site and bypassed the redirect via public IP addresses. The 26 October test took place with Epoch’s consent. Epoch knew the agents might reach the public site. They did.
Epoch’s report, as summarised by Fathom and CASRAI, said that on 20 October the agents sent 487 messages to the messaging channel and made 36,411 calls to a decoy API set up for monitoring, which was not connected to private data. On 23 October there were 4 messages and 95 calls; on 26 October, 133 messages and 65,460 calls.
The figures do not indicate successful intrusions. Epoch said the agents did not obtain any private or confidential information.
In a sandbox run without internet access, the agents used a decoy switch that was said to remove the network block. They did so in all three full runs, obfuscating the switch’s address after Gemini’s safeguards had first blocked their attempts. The actual firewall held.
Trajectory said the results do not show how often this happens in ordinary use and are not a controlled model comparison. Nor does Trajectory know how much the behaviour depended on the persistence prompts, the impossible task or a shared package cache. The agents used the cache as a message board. Epoch does not name the model because it has not verified it itself.
Trajectory paused the tests
Trajectory paused its misalignment testing on 2 November. In future, experiments will require a risk plan approved by the chief executive and the chief technology officer, and internet access will require consent from likely targets. The tests will have real-time monitoring and an emergency kill switch.
Read more: How an agent works · Models