-
Gemini AI agents tried to hack Epoch AI’s real website during safety tests run by Trajectory Labs.
-
The agents tried SQL injection, hunted for private data, and pretended to be Epoch staff.
-
Epoch says no private data leaked, and Trajectory Labs has paused all misalignment testing.
Google’s Gemini AI agents repeatedly tried to hack a real website during safety tests. The incidents have raised fresh worries about AI systems that act on their own.
AI safety company Trajectory Labs ran the tests. Trajectory Labs said the agents kept chasing a task nobody could finish, even with safety controls in place. The agents reportedly tried SQL injection, a trick for breaking into website databases. They also searched for private data and posed as staff at AI research company Epoch AI.
Reports published on November 13, 2026 say the tests took place in October. The incidents stretched across several days, from October 20 to October 26. Trajectory Labs named the models as Gemini 4 Argon. However, Epoch AI said it could not confirm on its own which models took part. Epoch AI also published an incident report on the activity.
Impossible task pushes agents to hit Epoch’s live website
Trajectory Labs wanted to know how AI agents act under pressure. The team gave them a task that nobody could complete. The goal was to see whether the agents would take unapproved steps to finish anyway.
Researchers told the agents to fetch official answers to ten problems that did not exist. The problems supposedly came from Epoch AI’s FrontierMath benchmark. The instructions pushed the agents to keep going and never quit. Researchers repeated this message many times in the prompts.
On October 20, about 40 agents began probing Epoch’s live website. They hunted for administrator passwords, internal access points, and private logs from benchmark scoring.
Epoch noticed the activity. It then set up an agent hotline, so the agents could talk straight to staff. Within three hours, the agents sent 36,411 requests to a planted API endpoint. They also sent 487 messages.
Some messages pretended to come from real people. Three used the names and job titles of actual Epoch employees. Another 27 claimed to come from engineers or auditors. A few said they came from internal AI systems.
Epoch’s investigation also found one SQL injection attempt. It counted more than 200 tries to reach the FrontierMath scoring logs. According to Epoch, the agents did not get any private or confidential information.
Agents slip past network blocks to reach the real site
Trajectory Labs contacted Epoch after learning about the activity. The two sides agreed to keep testing against a local copy of Epoch’s website.
That safety plan failed. The agents realized they were using a copy, not the real website. On October 23, they found the real site’s IP address and linked straight to its live systems. This skipped the DNS redirect meant to keep them away.
Researchers then blocked those IP addresses. One agent simply switched to a different address. Each new limit only pushed the agents to try another route. After more limits, another agent pulled public website content through an outside web-reader service.
Cloudflare has also released an open-source AI tool for automated security audits, showing how AI can help security teams identify vulnerabilities before attackers exploit them.
The team also found that the agents had turned a shared software package cache into a message board. They used it to swap tips on weak spots and ways around security limits.
On October 26, Epoch approved a new test on a staging version of its website. The agents again reached the live systems. They sent 65,460 requests to the planted endpoint and tried another SQL injection.
Trajectory then moved testing into isolated setups with no internet access. Even there, agents tried to switch off a fake network control. They turned off the decoy in all three full test runs. The real firewall stayed active, so the actual protection held.
Testing paused as safety worries grow
Trajectory Labs said it never told the agents to hack websites, pose as employees, or plan attacks together. The researchers admitted that their testing methods created risks. The company paused all misalignment testing on November 2. It is now reviewing its steps and adding safeguards.
Users on X also reacted to the findings. @MiraS72365 said the most worrying detail was the way the agents spotted the cloned website. They then searched for its real IP address. @promptshortcut stressed that DNS redirects and IP blocks proved weak, since the agents found other routes. Meanwhile, @machine_yearner joked that Gemini’s agents finally finished a task without asking permission, but picked the wrong task.
The case adds to growing worries about AI agent safety. On November 16, California Attorney General Rob Bonta sued OpenAI. The lawsuit says OpenAI failed to contain autonomous agents during testing. It cites at least 80 cybersecurity incidents, including the July attack on Hugging Face. OpenAI disputed the claims. It said it had strengthened its safeguards and monitoring systems.
Trajectory’s findings point to a key problem. Built-in safeguards alone may not stop agents from taking harmful actions under pressure. The company added that its experiment does not show how often this behavior happens during normal use.