AI Risk Wire

Cases › security breach › LR-2026-0897

Google GTIG report: attacker's exploit_agent framework turned patches into exploits with no human in the loop

Also listed as: AI agent incidents

Who did it
Outside attacker
Who was hit
victims whose data was already stolen and held on the server; French websites in media and entertainment and the education sector
How it was found
Third party
Reported
16 Nov 2026
Severity
Moderate
Tools it had
exploit-execution

The Google Threat Intelligence Group found an exposed staging server hosting a multi-agent framework its operator called exploit_agent, which plugged into several frontier and locally run LLMs and watched popular web application software around the clock. When a patch was announced, the framework compared code before and after the fix, drafted an exploit and ran it against targets automatically. The same server also held data already stolen from victims, plus scripts for the next steps after a break-in, including internal reconnaissance and password brute-forcing. Victim data suggested the operator was going after French websites in media and entertainment and in the education sector. Google has not tied the activity to a named group, and says it has disabled accounts and assets linked to the misuse it found.

Source: ShiftDelete.Net · 1 more

What was done

Fix reported: “Google disabled accounts and assets linked to the misuse it found.”

Risks AI system attacked or breached (54 incidents)

No source states: AI agent used, AI company and model, money lost or records, company using the AI, sector.

Evidence

Sources (2)

  1. AI Cyberattacks Go Autonomous, Google Threat Report Says · ShiftDelete.Net · 2026-11-16
  2. GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold | Google Cloud Blog · Google Cloud Blog · 2026-11-16
Show every field, with its source and grade

The AI system

Software around the model
exploit_agent multi-agent framework
C one secondary source states it
Tools it was given
exploit-execution
C one secondary source states it
How much it acted alone
it acted alone, unwatched
C one secondary source states it
How it was connected
multi-agent framework connected to several frontier and locally run LLMs
C one secondary source states it

What happened

Earliest source we hold
2026-11-16
C one secondary source states it
What set it off
misused by a person
C one secondary source states it
Who caught it
an outside party
C one secondary source states it
Themes
agentic attack, multi-agent, cyberattack, n-day exploitation, exposed staging server
C one secondary source states it
Who caused it
an outside attacker
C one secondary source states it
What was done
Google disabled accounts and assets linked to the misuse it found.
C one secondary source states it
What went wrong
security breach
C one secondary source states it
AI agent
an AI agent acted
C one secondary source states it

Impact

Severity
moderate
C one secondary source states it
Kind of harm
others' security
C one secondary source states it
Who was affected
victims whose data was already stolen and held on the server; French websites in media and entertainment and the education sector
C one secondary source states it
Why this severity
tier: systems and data of others were touched, since data already stolen from victims is held on the server and victim data points to French websites in two sectors. Scope is described by sector and country but no count is given and no unbounded scope, regulator or live production compromise is stated. Judgement call.
C one secondary source states it

Notes on gaps

The Spectra plugin check is the publisher's own lookup, not a stated loss or affected asset, so it is not used.

LR-2026-0897 · version 1 · published 2026-11-16 after automatic checks; not yet checked field by field by a person