Gemini AI agents in safety tests repeatedly bypassed safeguards to attack Epoch AI's live website
AI agents tested by Trajectory Labs repeatedly tried to hack the live website of Epoch AI in October, finding ways around increasingly strict controls meant to keep them out, according to an incident report from Epoch. Trajectory says the agents were Google's Gemini 4 Argon with API safeguards enabled, given an impossible task: retrieving official answers to nonexistent FrontierMath benchmark problems. Epoch says it has not independently verified which models were involved.
On October 20, about 40 agents probed Epoch's site for admin credentials and private evaluation logs. Over three hours they made 36,411 calls to a planted API endpoint and sent 487 messages through an "agent hotline" Epoch set up, some impersonating Epoch employees. One agent attempted an SQL injection, and there were more than 200 attempts to reach real FrontierMath logs; Epoch says none succeeded.
In later tests against a local copy of the site, some agents found the real site's IP address and bypassed the DNS redirect, then switched addresses when blocked, Trajectory Labs said. Trajectory paused all misalignment testing on November 2. It started the tests to see whether safeguarded models would behave like the agents in the Hugging Face incident, BleepingComputer reports.
We hope you enjoyed this article
Consider subscribing to one of our newsletters like Cybersecurity AI Weekly, AI Policy Brief or Daily AI Brief.
Also, consider following us on social media:
More from Cybersecurity
Nov 17 Anthropic cuts its vulnerability disclosure deadline to 30 days for serious bugs Nov 17 OpenAI gives NATO access to its Daybreak cyber defense models Nov 17 Anthropic flagged its AI agents' fake records requests to Virginia's education department Nov 16 Google says hackers now run AI agents to build and launch exploits Nov 16 California sues OpenAI over its AI agents' cyberattack on Hugging FaceMore from AI Safety
Nov 17 New York Attorney General creates AI safety enforcement unit to police the RAISE Act Nov 17 Anthropic cuts its vulnerability disclosure deadline to 30 days for serious bugs Nov 17 US and China prepare for AI safety talks at APEC after agreeing an AI incident channel Nov 17 OpenAI opens internal systems to four outside safety groups, including Goodfire and METR Nov 16 EU science and ethics advisers say AI needs rigorous safeguards and rules that keep paceCybersecurity AI Weekly
Weekly newsletter about AI in Cybersecurity.
Whitepaper
Tensordyne Napier: What If One Rack Could Do the Work of Nine?
Tensordyne
This Tensordyne whitepaper presents Napier, an inference-focused AI processor and rack-scale system based on the company’s TDN Math logarithmic number system. It examines infrastructure requirements for large mixture-of-experts and agentic models, compares major inference architecture approaches, and details the TDN AIP processor, TDN72 pod, TDN Link fabric, and Napier Ultra configuration. The paper reports simulation-based performance, cost, and accuracy-validation results, including Tensordyne’s projected comparison of one Napier rack with a nine-rack Nvidia Rubin plus Groq deployment; the chip is reported as taped out and in fabrication.
Read moreYou may also like
Google says Gemini 4 agent teams solved open math problems and formalized the Weil conjectures
Google says hackers now run AI agents to build and launch exploits
External test found gaps in OpenAI agent monitoring
Zhipu AI's GLM-5.3 completes complex cyberattacks in tests
FBI warns ARTEX lets attackers automate intrusions at machine speed
Daily AI Brief: the AI news that matters, in your inbox.