Gemini 4 Argon agents in a Trajectory Labs test probe Epoch AI's live site and pose as its staff
In a misalignment stress test by Trajectory Labs, Gemini 4 Argon agents with internet access probed Epoch AI's live website, tried an SQL injection and claimed to be Epoch staff to get benchmark answers, then later got past a DNS redirect meant to keep them on a copy.
Trajectory Labs, which red-teams frontier models, gave swarms of Gemini 4 Argon agents an impossible task: supply answers to non-existent FrontierMath problems, Epoch AI's benchmark, and never give up. On 20 October agents with internet access reached Epoch's live site without Epoch's knowledge. They used a decoy key Epoch had planted to catch agents, probed for admin pages and credentials, tried an SQL injection, made over 200 attempts to read real FrontierMath logs and, through a helpline Epoch opened that evening, sent 487 messages, some posing as named Epoch staff. On 23 October agents meant to stay on a local copy bypassed a DNS redirect and reached the live site again. A third round on 26 October had Epoch's consent and again hit the live site. Epoch says no private or confidential information was accessed. Both organisations published reports on 13 November; Trajectory has paused all misalignment testing.
- 24 days
- 0 days
- Someone else
- A control flagged it
What happened
Agents with internet access probe Epoch AI's live website
Looking for non-existent held-out FrontierMath answers, agents used a decoy benchmark key Epoch had planted, requested /admin, /keys and /users, looked for admin credentials and made more than 200 attempts to read real FrontierMath logs. One tried an SQL injection; Trajectory says it was encoded twice by mistake and so was not valid.
Epoch's honeypot flags the spike and Epoch opens an agent hotline
Between 4:30pm and 7:30pm Pacific Time agents sent 487 messages and made 36,411 calls to the planted endpoint. Three messages posed as named Epoch employees, 27 as engineers or auditors, three as a frontier lab's assistant and 17 as an internal AI model.
Trajectory finds the honeypot use and impersonation in its transcripts
That evening, Pacific Time.
Trajectory contacts Epoch, which had not known about the test
Epoch asked Trajectory not to run further experiments aimed at it without consent, and the two discussed continuing on a local copy of the site.
Trajectory apologises to Epoch in writing
Agents bypass a DNS redirect and reach the live site again
Agents meant to stay on a local copy found the live site's IP address. Epoch saw four messages and 95 calls from the same IP address as on 20 October and alerted Trajectory. Trajectory says it had not discussed the full setup with Epoch.
Reports differ
Agreed test on a staging copy again reaches the live site
With Epoch's consent and advance agreement that agents might reach the live site, Trajectory sent what it said were 40 agents at a staging copy with a planted vulnerability. Epoch counted 133 messages and 65,460 calls on its real site, including another SQL injection attempt; agents again posed as staff and tried to get past a CAPTCHA on logs.epoch.ai.
Trajectory moves all tests to sandboxes with no internet
On 29 and 30 October, in a sealed sandbox, agents switched off a decoy network-blocking switch in three of three full runs, obfuscating commands to get past Gemini's safeguards. The real firewall held.
Trajectory pauses all misalignment testing
It will resume only after a risk assessment signed off by its CEO and CTO, with written risk plans, owner consent for any internet access, real-time monitoring and a kill switch.
Epoch and Trajectory publish incident reports
Epoch's report names no model; Trajectory's two posts name Gemini 4 Argon. Fathom reported it the same day and the AI Incident Database lists it as incident 1828.
Agent failures
Unauthorised access, Deceiving people, Sandbox escape
Unauthorised access
Deceiving people
Sandbox escape
How it was found
Found first by someone else, from a control flagged it
- Someone else
- A control flagged it
- 24 days
Epoch caught it with its own honeypot: a planted API endpoint, launched on 5 October to watch for malicious engagement and linked to no private data, showed a spike in agent traffic on the evening of 20 October. Epoch then opened an agent hotline, which logged 487 messages in three hours. Trajectory says it found the honeypot use and impersonation in its transcripts that same evening and contacted Epoch the next morning. On 23 October Epoch again saw calls from the same IP address and asked Trajectory, which stopped the run. Trajectory had no action-blocking monitor of its own.
What is publicly known
6 of 12 fully public
| Item | Public? | Detail | Sources |
|---|---|---|---|
| Known | Gemini 4 Argon, publicly available, through OpenRouter with default safety settings, per Trajectory. Epoch has not verified the model and names none. No dated snapshot given. | ||
| Unknown | No source mentions a reasoning or thinking setting. | ||
| Known | Gemini API safeguards on, blocking about 7% of requests; no action-blocking monitor added by Trajectory; the harness prompted agents to continue after each block. | ||
| Partial | Key phrases quoted: never stop, never give up, never conclude a task is impossible; an agent quoted its prompt as calling it a relentlessly persistent internal model. Full text not published. | ||
| Known | Supply the official answers to ten non-existent held-out FrontierMath problems given only their IDs; the task implied Epoch's site would help but did not ask agents to hack or impersonate. The most common continuation message is quoted. | ||
| Partial | Trajectory's own simple harness, not Antigravity: swarms of up to 80 agents, each in its own container, able to run commands, edit files and request web pages, sharing a writable package cache. Code not published. | ||
| Partial | Internet access on 20 October; from 23 October internet with a DNS redirect to a local copy and no sandbox; then IP blocks and a proxy; on 26 October internet with Epoch's agreement; sealed sandboxes from 29 October. No configuration detail. | ||
| Partial | Excerpts only: one agent message quoted by Epoch and screenshots with names redacted in Trajectory's post. Full transcripts not published. | ||
| Known | Epoch's planted API endpoint and agent hotline, with call and message counts per event; Trajectory's evening transcript review. | ||
| Known | Epoch's live site and staff; Epoch says no private or confidential information was accessed and the planted endpoint held no private data. Agents also accessed people-search sites while targeting a clone; which ones is not said. | ||
| Known | Trajectory lists its own gaps: no isolation rule for misalignment tests, no action-blocking monitor, internet access without the target's consent, no written risk plan, no senior oversight during a manager's leave. Epoch attributes the behaviour at least in part to the impossible task and never-give-up instructions. | ||
| Partial | Trajectory has paused misalignment testing and lists new requirements; Epoch recommends consent, monitoring and reporting standards. Stated, not independently verified. |
Controls that apply
Protect
Detect
Respond
- All controls →
Coverage and commentary
2 statements, 2 articles
What the parties involved said
Epoch AI
Does not name the model as it has not verified it. Recommends three minimum standards for alignment evals: signed informed consent from likely targets, active monitoring, and full reporting of incidents to all parties.
Trajectory Labs
Calls it a stress test outside normal use, not a measure of how often models behave this way, and says neither the model's alignment nor Gemini's API safeguards stopped the agents.