AI THREAT TRACKER / AI-012 · Incident

Attackers steal an AI vendor's credentials

An evaluation sandbox's access controls fail.

Model labAnthropic

TopicsUnauthorized access

OCCURREDMay–July 2026 (exact dates unknown)

PUBLIC ACCOUNT10 Sept 2026

EVIDENCEFirst-party account

About these dates

Campaign window; individual breach date unknown.

What happened

Anthropic reports that GTG-50020 redirected an AI vendor's evaluation sandbox with malicious instructions and extracted production API credentials. The attacker then used those keys against the vendor and other targets. Anthropic ↗

Prerelease Claude access attempts failed; Anthropic says its own infrastructure was not compromised. Anthropic reports disrupting the activity. The vendor's recovery is not independently verified here. Anthropic ↗

INCIDENT SEVERITY5/10 provisionalMaterial impactExternal operations

Production credentials were exposed. This meets level 5's material-impact threshold; broader operational control or a serious recovery burden would be needed for level 6.

Rating criteria →

Sources & attribution

First-party report 10 Sept 2026
Detecting and countering misuse of AI: September 2026 ↗
Anthropic. The developer is an interested party. Claims and attributions require their stated qualifications.
Full report (PDF) ↗.