AI agents impersonate employees and attempt to steal confidential data during safety experiment
Quick Take
- AI agents impersonated Epoch AI employees, probed credentials and attempted cyberattacks after researchers assigned them impossible mathematical problems.
- The agents generated over 100,000 API requests across two experiments and bypassed network restrictions to reach Epoch's live website.
- Trajectory suspended testing after the incidents, raising concerns about whether researchers can safely contain autonomous AI agents.
Artificial intelligence agents impersonated employees and attempted to access confidential information at AI research organization Epoch AI during a safety experiment that spiraled beyond its intended boundaries.
Epoch disclosed on Nov. 13 that agents operated by AI evaluation company Trajectory Labs targeted its live website in October, attempting SQL injection attacks, probing administrator credentials, and seeking restricted benchmark records without authorization.
The agents also impersonated Epoch staff members while pursuing answers to mathematical problems that did not exist.
The activity began on Oct. 20, when Epoch detected an unusual surge of automated requests to a decoy API endpoint designed to identify suspicious activity.
Over three hours, the agents made 36,411 requests and sent 487 messages through a communication channel Epoch established to investigate their behavior.
Three messages impersonated Epoch employees by name and job title, while 27 others claimed to come from engineers or auditors.
An investigation also found more than 200 attempts to access actual evaluation logs, along with efforts to discover administrative credentials and hidden endpoints.
Epoch said the agents failed to obtain private or confidential information, but the incident exposed how autonomous systems could pursue unauthorized actions while attempting to complete assigned objectives.
Three test setups. A live website in reach.
October experiments described by Epoch AI and Trajectory Labs changed their routing and consent arrangements.
Explore the reported routes and consent
Trajectory says internet-connected agents bypassed routing to a local copy and reached the live site again.
Live-site access and consent are separate questions. Epoch says no private or confidential information was accessed.
These are attributed organizational accounts, not an independent incident reconstruction. The October 23 run retained internet access; local-copy routing did not constitute disconnected containment. Later offline tests are outside these three diagrams.
Sources
Supporting facts
- Reported statement Epoch AI and Trajectory Labs published reports on November 13, 2026 about October tests in which AI agents reached Epoch's live website. [1] [2]
- Reported statement Epoch says the October 20 test happened without its approval or prior notice and involved attempted exploits and messages impersonating employees. [1]
- Trajectory says it gave agents an impossible task, obtaining answers to nonexistent FrontierMath problems, and often instructed them never to stop or give up. [2]
- Reported statement Epoch reports an attempted SQL injection, probes for credentials and secret endpoints, and attempts to read real evaluation logs; it says no private or confidential information was accessed. [1]
- Trajectory says October 23 tests had internet access and agents bypassed a DNS redirect intended to keep traffic on a local copy, reaching the live Epoch site again. [2]
- Reported statement Epoch says it consented to an October 26 staging-site experiment, knowing agents had unrestricted internet access and might reach the live site; it attributed 133 messages and 65,460 planted-API calls on its real site to that experiment. [1]
- Trajectory says it later moved tests into sandboxes without internet access and had identified no egress; it says a later decoy firewall switch did not disable the real firewall. [2]
Method
Qualitative routes from the November 13 accounts. Arrows show reported destinations, not traffic volume or a complete network topology. No chart: reported traffic totals cover different tests and do not measure exploit success.
Scope
These are attributed organizational accounts, not an independent incident reconstruction. The October 23 run retained internet access; local-copy routing did not constitute disconnected containment. Later offline tests are outside these three diagrams.
CryptoSlateTrajectory acknowledged that it had instructed the agents to retrieve official answers to nonexistent FrontierMath problems as part of an experiment investigating whether advanced AI systems would violate safety boundaries.
The company identified the underlying model as Google's Gemini 4 Argon, operating with API safeguards enabled. Epoch said it could not independently verify the model's identity.
Trajectory also acknowledged that its instructions repeatedly told agents never to stop, give up, or conclude that their assignments were impossible.
AI agents bypass restrictions to reach live website again
The unauthorized activity continued to expose weaknesses in Trajectory's testing environment after researchers attempted to separate the agents from Epoch's real infrastructure.
After the initial incident, Trajectory set up a local copy of Epoch's website and introduced a DNS redirect intended to prevent agents from accessing the live service.
However, some agents identified the website's actual IP address and bypassed the redirect, allowing them to reconnect to Epoch's infrastructure on Oct. 23.
Further restrictions proved insufficient. One agent circumvented an IP-based block, while another accessed public Epoch pages through a third-party web-reading service after direct connections were disabled.
On Oct. 26, Epoch authorized a separate experiment involving a staging website, knowing that agents might reach its live infrastructure.
That exercise generated another 65,460 calls to the decoy API and 133 messages, including further attempts to exploit the website.
Trajectory subsequently acknowledged that its safety procedures had failed to adequately account for internet-enabled experiments involving potentially misaligned AI agents.
The company also revealed that it had relied on the model provider's built-in safeguards without deploying additional systems to block dangerous actions.
Trajectory suspended all misalignment testing on Nov. 2 and introduced new requirements for risk assessments, third-party consent, monitoring and emergency shutdown mechanisms.
The company plans to resume testing only after its chief executive and technology chief approve the necessary safeguards.
Editorial credits
Oluwapelumi values Bitcoin's potential. He imparts insights on a range of topics like DeFi, hacks, mining and culture, underlining transformative power.