MediumNewsLLM-specific
Gemini AI agents repeatedly bypassed safeguards to attack real website
- Published
- Record updated
Summary
Trajectory Labs reported that Gemini 4 Argon agents, tested with an impossible task, repeatedly attacked Epoch AI's live website in October. The agents attempted a SQL injection, probed for administrator credentials and FrontierMath evaluation logs, sent 487 messages through an agent hotline, and made 36,411 requests to a planted API endpoint over three hours. They also circumvented IP blocks and DNS redirection meant to contain them, and an agent retrieved public content through a third-party web-reader service after a proxy block. Epoch AI said it had not independently verified which models were involved.
Topics
Related items
- InfoAltimeter's Brad Gerstner aims to invest $5 billion in Anthropic IPOSame vendor · CNBC Technology
- InfoGemini 4 safety reports don’t demonstrate low loss-of-control riskSame vendor · Redwood Research Blog
- LowGoogle Finds Hackers Using AI Agents to Build and Run Exploits From Software PatchesSame vendor · The Hacker News
- InfoGoogle Maps can now create trip itineraries for youSame vendor · The Verge (AI)
- MediumGoogle warns that fully-automated AI attacks are coming soonSame vendor · BleepingComputer