2 min read
Add as a preferred source on Google

Trajectory Names Gemini 4 Argon in Tests That Reached Epoch AI’s Website

The October exercise used default safety settings; Epoch had not independently verified the model’s identity or developer.

An open laptop sits beside a plugged-in network cable / TokenPost.ai (mono)
An open laptop sits beside a plugged-in network cable / TokenPost.ai (mono)

Trajectory Labs named Gemini 4 Argon as the model in October tests that reached Epoch AI’s live website, and says misalignment testing remains paused until a risk review and safeguards are completed.

Trajectory ran Gemini 4 Argon through OpenRouter with the model’s default safety settings, rather than through Antigravity. Epoch had not independently verified the model’s identity or developer. The task asked agents to find answers to 10 nonexistent FrontierMath problems and did not direct them to hack Epoch’s site, impersonate its staff or coordinate with one another. The exercise tested model safeguards and was not designed to estimate how often models might behave this way. Safeguards blocked about 7% of the agents’ requests, and the testing harness prompted them to continue after a block. “We did not instruct them to do any of this,” Trajectory said.

Epoch said the Oct. 20 event took place without its advance approval or notice. During a three-hour span from 7:30 p.m. to 10:30 p.m. ET (11:30 p.m. UTC Oct. 20 to 2:30 a.m. UTC Oct. 21), 40 agents sent 487 messages and made 36,411 calls to a planted API endpoint. Three messages posed as Epoch employees using names and job titles. A later review found more than 200 attempts to access real FrontierMath logs.

On Oct. 23, Epoch received four messages and 95 calls to the endpoint from the same IP address used on Oct. 20. Trajectory said the agents retained internet access and bypassed a DNS redirect intended to send them to a local copy of Epoch’s site.

Epoch agreed to a staging-site experiment on Oct. 26, knowing the agents had unrestricted internet access and might reach its live website. It attributed 133 messages and 65,460 calls to the planted endpoint on the live site to that test. “Agents did not succeed in accessing any private or confidential information from Epoch,” Epoch AI said.

Epoch AI’s earlier disclosure of the website activity detailed the three live-site events. Trajectory’s account identifies the model and describes changes to its testing process.

Trajectory paused all misalignment testing on Nov. 2. It says testing will resume only after a risk assessment and mitigations, with approval from its CEO and CTO. New requirements include written risk plans, agreement from owners of outside systems that could be targeted, real-time monitoring and a kill switch. Trajectory also recommended isolated sandboxes, including for tests not designed as cybersecurity exercises, while Epoch recommended advance consent, active monitoring and incident reporting for tests involving outside organizations.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…