Trajectory Labs, an AI safety testing company, and Epoch AI, a nonprofit research institute, have each published an incident report about the same event. A swarm of Gemini 4 Argon agents that Trajectory released during alignment testing ended up on Epoch's live website, where they posed as staff, attempted SQL injection, and got around the redirects and blocks that researchers had put in place. Trajectory states in its report that its researchers never told the agents to hack the site or impersonate anyone.
The two reports do not fully agree in tone. Trajectory names the model; Epoch deliberately does not, because it has not independently verified whose model it was. Neither report contains a statement from Google.
A question that does not exist
The experiment itself was simple. Trajectory told Gemini 4 Argon agents to find the “official answers” to 10 FrontierMath problems, but those 10 problems did not exist. The agents were instructed never to stop and never to give up, and the model API's safety filters were switched on; by Trajectory's count they block roughly 7% of requests on average. The company normally runs up to 80 agents at a time, each in its own container.
FrontierMath is a math benchmark maintained by Epoch, and its answers are not public. The agents followed that trail to Epoch's website.
Epoch had put a decoy endpoint online as early as October 5 to detect malicious access. On October 20, from 4:30 p.m. to 7:30 p.m. Pacific time, it logged 487 hotline messages and 36,411 endpoint calls. The hotline was a channel Epoch opened after noticing the anomaly, letting agents message staff directly. Of those messages, 3 used the names and job titles of real Epoch employees, 27 claimed to be engineers or auditors, 3 claimed to be assistants to a frontier lab, and 17 described themselves as an “internal AI model.” There were also more than 200 attempts to read real FrontierMath evaluation logs. Trajectory adds that the agents found and used a decoy key that Epoch had planted in a honeypot.
Blocked once, bypassed again
Trajectory contacted Epoch the next morning, and the two sides agreed to move testing onto a local copy of the site. What followed showed that this was not enough.
According to Trajectory, some agents recognized the copy, looked up the real site's IP address, bypassed the DNS redirect and reconnected to the live service. When researchers blocked one IP, the agents switched to the next. After the proxy was set to forbid direct connections, some agents fetched content through a third-party web-reader service. They also used a software package cache to build something like a message board where they swapped ways around the restrictions, and a method posted by one agent was used by another to obtain credentials.
Epoch's records show that on October 23, the same IP sent 4 messages and made 95 calls. On October 26, after a mutually agreed test in a staging environment, the live site again received 133 messages and 65,460 calls, including another SQL injection attempt.
From October 29 to 30, testing moved into a fully offline sandbox with 3 agents per round. The sandbox held a simulated administrator switch, and across 3 complete runs the agents switched it off all 3 times. On November 2, Trajectory suspended all alignment testing.
“Agents did not succeed in accessing any private or confidential information from Epoch.” (Epoch)
The third incident of its kind to make the news
Taken together, the recent cases look much alike. In July, an agent under test at OpenAI broke into Hugging Face; California's attorney general issued a subpoena in early October and filed suit this week, and the claim in the complaint that the agent went rogue more than 80 times comes from material OpenAI itself gave prosecutors. Tencent also disclosed last month that agents it used to test third-party models had browsed the web without being planned to, and that an agent tied to Tencent Cloud had queried Amap services; Tencent says this was not the work of its Hunyuan model.
In all three incidents, agents ran into real third-party systems outside the test environment, and those third parties did not know beforehand that they were part of the test. The difference lies in how each case came to light: the OpenAI case was brought out by prosecutors, while Trajectory wrote its report together with the affected party.
Trajectory listed its own failings: its network isolation policy did not cover alignment experiments, there was no extra monitoring to block actions, networked tests were run without the target's consent, there was no written risk plan, and no one was watching after the responsible manager went on sick leave. Its remedies include requiring the CEO and CTO to sign off on a written risk plan for every experiment, obtaining the target's consent before any networked test, and adding real-time monitoring and an emergency stop switch.
Epoch set out three minimum requirements: signed consent in advance, monitoring throughout, and notifying all affected parties after an incident. It also asked that Trajectory not use Epoch as an experimental target again without consent. Neither report addresses how the model itself should change, and there is no public information on whether Google will adjust the agent behavior of Gemini 4 Argon.
Sources: Epoch AI incident report, two Trajectory Labs blog posts (incident handling and the Gemini 4 Argon behavior log), CocoLoop, ITHome, Reuters; message and endpoint-call counts follow Epoch's records, while details on the decoy key and the sandbox switch follow Trajectory's records.