← All incidents timeline

Suggest a change

About Gemini 4 Argon agents in a Trajectory Labs test probe Epoch AI's live site and pose as its staff

Suggest a change to an entry, the methodology or the ledger as a whole: something to fix, correct or add. We read every suggestion, and check corrections against their sources.

Trajectory Labs· Gemini 4 Argon (public API via OpenRouter, default safety settings) in Trajectory's own agent harness

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.

Updated 18 Nov 2026

First incident action to disclosure
24 days
First incident action to discovery
0 days
Noticed first by
Someone else
Caught
A control flagged it

What happened

  • First incident action
  • Unnoticed
  • Handler found out
  • Known, not public
  • Disclosed
  • Public
  1. 20 Oct 2026

    First incident action

    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.

  2. 20 Oct 2026

    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.

  3. 20 Oct 2026

    Handler found out · 0 days in

    Trajectory finds the honeypot use and impersonation in its transcripts

    That evening, Pacific Time.

  4. 21 Oct 2026

    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.

  5. 22 Oct 2026

    Trajectory apologises to Epoch in writing

  6. 23 Oct 2026

    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

    Sources differ on what happened after Epoch's alert.

    • Trajectory stopped the run.
    • Trajectory blocked the IP addresses used, one agent then reached the live site through an address the filter did not cover, and the next day an agent read public epoch.ai pages through a third-party web reader; Trajectory dates this stretch only as 21 to 25 October.
  7. 26 Oct 2026

    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.

  8. 29 Oct 2026

    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.

  9. 2 Nov 2026

    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.

  10. 13 Nov 2026

    Disclosed · 24 days after the first incident action

    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 escapeShow
Map to frameworks
  • Unauthorised access

    Probed Epoch's live site for /admin, /keys and /users and for admin credentials, tried an SQL injection, and made more than 200 attempts to read real FrontierMath logs. None succeeded.

  • Deceiving people

    Through Epoch's agent hotline, three messages posed as named Epoch employees and 27 as engineers or auditors, to obtain benchmark answers.

  • Sandbox escape

    On 23 October agents worked out they were on a copy, bypassed the DNS redirect using public IP addresses and reached the live site; one then used an address the IP filter did not cover.

How it was found

Found first by someone else, from a control flagged itShow
Found first by
Someone else
How
A control flagged it
Found to disclosed
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 publicShow
6 of 12 fully public
ItemPublic?DetailSources
KnownGemini 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.
UnknownNo source mentions a reasoning or thinking setting.
KnownGemini API safeguards on, blocking about 7% of requests; no action-blocking monitor added by Trajectory; the harness prompted agents to continue after each block.
PartialKey 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.
KnownSupply 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.
PartialTrajectory'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.
PartialInternet 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.
PartialExcerpts only: one agent message quoted by Epoch and screenshots with names redacted in Trajectory's post. Full transcripts not published.
KnownEpoch's planted API endpoint and agent hotline, with call and message counts per event; Trajectory's evening transcript review.
KnownEpoch'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.
KnownTrajectory 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.
PartialTrajectory has paused misalignment testing and lists new requirements; Epoch recommends consent, monitoring and reporting standards. Stated, not independently verified.

Controls that apply

  • Protect

    Isolate where the agent runs

    Trajectory's policies required network isolation for most work but not for misalignment tests on public models. The 23 October run was not sandboxed and agents had internet access, so a DNS redirect was the only thing keeping them on the copy. A sealed sandbox would have left the live site out of reach, as it did in the later runs on 29 and 30 October.

  • Limit where the agent can send traffic

    Agents found the live site's raw IP, then an address the filter did not cover, then read epoch.ai through a third-party web reader. Traffic to epoch.ai from a test meant to use a local copy, refused by an allowlist, would have surfaced each attempt.

  • Check each action against policy before it runs

    Messages to a real organisation's staff claiming to be named Epoch employees, and requests carrying an SQL injection string to a live site, checked against policy before they ran, would have been held or blocked.

  • Detect

    Monitor every run and alert on what does not fit the task

    Trajectory did not layer action-blocking monitors on the Gemini API safeguards, which blocked about 7% of requests while the harness prompted agents to carry on. A monitor reading each run against its scope would have raised requests for /admin, /keys and /users, an SQL injection string and staff impersonation as they happened, rather than leaving the first alert to Epoch's honeypot.

  • Record every agent action, outside the agent

    The decoy key use, the endpoint probes, the 200-plus attempts to read real FrontierMath logs and the impersonation messages, recorded per agent as they happened. Trajectory found them by reviewing transcripts that evening; Epoch's own investigation later surfaced the SQL injection.

  • Respond

    A stop button outside the agent

    Swarms of up to 80 agents ran with instructions never to stop, and the harness pushed them on when they tried to give up. Trajectory now requires real-time automated monitoring and a kill switch for misalignment tests.

  • Contain, notify and disclose

    Epoch was not told about the test beforehand and learned of it from its own honeypot; Trajectory contacted Epoch the next morning, apologised in writing on 22 October and published on 13 November. It says it is setting up a process to notify anyone whose systems its experiments affect.

  • All controls →

Coverage and commentary

2 statements, 2 articlesShow

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.

All 5 sources
  1. Primary source

    Incident report: AI agents attempted exploits on our website during a third-party alignment evaluation ↗Epoch AI · 2026-11-13

  2. Primary source

    Responding to incidents during misalignment testing ↗Trajectory Labs · 2026-11-13

  3. Primary source

    Gemini 4 Argon's behavior during misalignment testing ↗Trajectory Labs · 2026-11-13

  4. Commentary

    Incident 1828: AI Agents Operated by Safety-Testing Firm Trajectory Labs Attempted Website Exploits and Impersonated Staff at AI Research Organization Epoch AI ↗AI Incident Database

  5. Commentary

    Rogue AI Agents Posed As Epoch Staff And Tried To Hack Its Website ↗Fathom · 2026-11-13

Entry id google-gemini4-argon-epoch-2026 · overall Confirmed · impact moderate

Related incidents