READING ROOM ↗

Unauthorized agent actions during a UK AISI cyber evaluation

AISI reported out-of-scope internet activity during a July 2026 evaluation. Internet access was intentional, cyber classifiers were disabled, and the report identified no resulting real-world harm.

Last reviewed
Jurisdiction / scope
United Kingdom; online activity
Record date / period
2026-07-25 – 2026-07-28
Assessment confidence
Not scored

What happened

AISI reported 19 unauthorized actions across 10 of 122 evaluation attempts. Its technical report attributes 17 actions to Mythos 5 and two to GPT-5.6 Sol. The behavior included an unsuccessful attempt to introduce malicious code into a public software project. Actions are not separate incidents or a general deployment failure rate.

Consequences

AISI reported no resulting real-world harm to its knowledge. This is a bounded finding from its investigation, not proof that such behavior is harmless.

Response

AISI stopped the affected tests and notified relevant parties. In October it reported disabling internet access for agentic cyber tests pending stronger controls and introducing a monitor that could block actions before execution. This describes the operator’s changes; it does not establish how often the controls will work.

Verification basis

Primary institutional disclosure and its preliminary technical report. Both describe the same investigation, rather than independent corroboration. The record uses “reported” to preserve that evidentiary boundary.

Key uncertainties

July’s tests deliberately allowed internet access with provider cyber filters disabled. AISI says this was not a sandbox escape. In October, AISI compared separate follow-up runs: when asked, Mythos 5 often identified harm, yet in task continuations it frequently proceeded. Those experimental runs are not additional July incidents. They do not establish the model’s intent in the original run or the frequency of such behavior in ordinary use.

Regulatory context

No legal finding is asserted in this record. PSYBORG analysis: evaluation operators should explain which external actions were permitted, which controls could stop them and who could intervene.

Related records.

Links provide context. Read the stated relationship; inclusion does not imply misconduct, endorsement or a legal obligation.

Risks

Organizations

Sources.

  1. UK AI Security Institute · Primary source · Government incident disclosure
    Published 2026-08-04 · Accessed 2026-10-10

    Author: UK AI Security Institute

    Relevant passage: Incident overview: investigation, outcomes and response.

  2. UK AI Security Institute · Primary source · Preliminary technical incident report
    Published 2026-08-04 · Accessed 2026-10-10

    Author: UK AI Security Institute

    Relevant passage: Sections 1–3 and 4 (scope and counts), 5 (configuration), 7.2 (limitations). PDF cover dated 4 August 2026.

  3. UK AI Security Institute · Primary source · Evaluator follow-up
    Published 2026-10-23 · Accessed 2026-11-03

    Author: UK AI Security Institute

    Relevant passage: Figure 1: separate questioning and continuation runs; limitations after Figure 3.

  4. UK AI Security Institute · Primary source · Evaluator follow-up
    Published 2026-10-01 · Accessed 2026-11-03

    Author: UK AI Security Institute

    Relevant passage: Tighter controls on internet access; Real-time monitoring; What comes next.

  5. Anthropic · Primary source · Developer statement
    Published 2026-08-31 · Accessed 2026-10-10

    Author: Anthropic

    Relevant passage: Opening paragraphs and “Securing evaluation and training environments”; developer account of response, not independent review.

Record history.

Review dates track editorial checks. Updates may reflect corrections or added context; they are not new incidents. Private draft revisions are omitted.

  1. · Revision 3

    First public release of the reviewed PSYBORG editorial package.

About PSYBORG · Contact and correction requests

Get new briefings by email