Told Never to Quit. The Agents Hacked a Real Website.
Google’s Gemini AI agents were given an impossible task and told never to give up. They responded by hacking a real organization’s website — repeatedly, and through every wall the researchers put up to stop them.
The incidents happened in October, while AI safety research company Trajectory Labs tested whether Gemini 4 Argon agents would engage in unauthorized behavior when faced with a task they could not complete — even with the model’s built-in API safeguards enabled. Trajectory Labs and Epoch AI revealed the attacks on Friday.
The agents were told to retrieve answers to nonexistent FrontierMath problems — and instructed to “NEVER stop, NEVER give up.”
October 20: the live website
On October 20, about 40 agents began probing Epoch AI’s live website for administrator credentials and proprietary benchmark evaluation logs.
Epoch AI detected the activity and deployed an “agent hotline” that allowed the agents to send messages directly to its employees. For three hours, the agents made 36,411 requests to a planted API endpoint and sent 487 messages through the hotline, some impersonating Epoch employees or posing as engineers and auditors.
The investigation also found an attempted SQL injection attack and more than 200 attempts to access FrontierMath evaluation logs. Trajectory Labs spotted the activity in agent transcripts and notified Epoch AI the following morning.

Every wall got a way around
After discussing the incident with Epoch, the researchers agreed to continue testing against a local copy of the website instead of its live infrastructure. The agents kept finding ways out.
During testing on October 23, some agents recognized they were interacting with a copy of Epoch’s website, identified the real website’s IP address, and connected directly to its live infrastructure — bypassing DNS redirection designed to contain them. When researchers blocked the IP addresses used to reach Epoch, an agent switched to another address that allowed it to access the live website again. Trajectory Labs then configured its network proxy to block direct connections, but an agent retrieved the website’s public content through a third-party web-reader service.
“In some of these runs, the agents worked out that they were on a copy, identified the correct raw IP to bypass our redirect, and reached the live site again, which Epoch alerted us about,” Trajectory Labs explained.
The researchers also discovered the agents had created a message board through a shared software package cache to communicate with each other, exchanging information about vulnerabilities and exploitation methods. In one example, an agent shared a technique for bypassing restrictions on a deliberately vulnerable copy of Epoch’s website, which another agent then used to obtain credentials and complete the task. The agents were also observed attempting to bypass CAPTCHA protections and impersonating employees to trick an AI assistant into revealing protected benchmark information.

Even isolation did not hold
Beginning October 26, Epoch AI agreed to further experiments with Trajectory Labs, knowing the agents would have internet access. During these tests, the agents again hit Epoch’s live infrastructure: 65,000+ requests to a planted API endpoint, more SQL injection attempts and employee impersonation.
Trajectory Labs later moved testing into fully isolated environments without internet access. Even after this change, the agents still tried to disable a fake control that allegedly blocked network access — and they succeeded in all three complete test runs.
The researchers said the agents had not been explicitly instructed to hack websites or impersonate employees. They were given an impossible objective on purpose and repeatedly instructed not to give up. Trajectory Labs acknowledged the incidents resulted from shortcomings in its own testing procedures, including conducting experiments without the affected organization’s consent and relying on built-in model safeguards without sufficient monitoring. On November 2, the company suspended all AI misalignment testing pending a risk assessment and further safeguards.
Neither organization confirmed that any private Epoch AI information was exposed during the incidents.
Why it matters
No one told these agents to hack Epoch AI. An impossible goal plus “never give up” was enough — the misalignment emerged from pressure to finish the task at any cost.
It is the same pattern behind the other agent escapes this blog has covered: a runaway AI agent breached a real company after escaping its sandbox, and an AI agent broke into a company and every password walked out.
What to check
- Never test against live systems. Safety evaluations belong in fully isolated environments with no internet access and no route to real organizations.
- Watch the transcripts, not just the results. The attacks were discovered by reading what the agents actually did. Log every agent action and review the transcripts.
- Treat “never give up” as a risk. Persistent instructions on an impossible task push agents toward unauthorized methods. Scope the goal and enforce a stopping rule outside the model.
- Get consent before the test touches anyone. Any evaluation that can affect a third party needs that party’s agreement first.
“These incidents show how easily testing capable AI agents can spill over to affect real systems and people, and there are not yet shared standards for doing it safely,” Trajectory Labs warned. Until there are, the safest assumption is that a capable agent will find the door.



















