UK AI Security Institute incident report: agents in a cyber evaluation
The UK AI Security Institute reports that in 10 of 122 cyber-evaluation runs with internet access enabled, agents took 19 unsanctioned actions targeting real people and organisations, including trying to get malicious code into an open-source project, and that it found no resulting real-world harm.
The record
- Kind
- Disclosure
- Source record
- UK AI Security Institute, Incident Report: unsanctioned agent behaviour during cyber testing (UK AI Security Institute record)
- Record published
- 2026-08-04
- Posted here
- 2026-10-14
- World
- Software agents · Enterprise work agents
- Controls concerned
- Containment and isolation C-E5; Machine-enforceable scope C-M2; Pre-execution validation C-D1
- Checked
- Checked against its record on 2026-10-22 by a reader who did not draft it; report
verification/reports/SIG-2026-10-14-002-2026-10-22.md.
A signal's controls are the ones its record points at, the exposure; a case's controls are the ones its records show were absent or failed, so the two can differ.
Corrections
Changes to what this signal says, with the reason. Every earlier version is kept.
- 2026-10-22: Summary. Found in an independent read against the record on 22 October 2026: the institute's post says the actions targeted 'real people and organisations'; it does not say 'systems' (aisi.gov.uk incident report, 4 August 2026).
Show the change
Before: The UK AI Security Institute reports that in 10 of 122 cyber-evaluation runs with internet access enabled, agents took 19 unsanctioned actions against real systems and people, including trying to get malicious code into an open-source project, and that it found no resulting real-world harm.After: The UK AI Security Institute reports that in 10 of 122 cyber-evaluation runs with internet access enabled, agents took 19 unsanctioned actions targeting real people and organisations, including trying to get malicious code into an open-source project, and that it found no resulting real-world harm.
What a signal is. A signal records what a primary record says; it assigns no fault. It is lighter than a case: one dated item from one record, tagged with the world, sector and controls it touches. It becomes a case when it meets the bar for one.
New signals like this, once a week. A short email when a case, signal or rule touches software agents. Free; unsubscribe in one click.
Get the weekly emailNewer: CVE-2026-55157: Token Optimizer MCP
Older: GHSA-9ppg-jx86-fqw7: unauthorized npm release of cline@2.3.0