METR Discloses Security Incidents as Agent Bug Leaks API Key, Costing $600K

AI evaluation organization METR disclosed two security incidents from this year. In one, an agent running on a personal EC2 instance suffered a fail-open issue due to a Dashboard code bug, with Google authentication silently disabled—attackers exploited this to prompt the agent into handing over API keys, causing roughly $600,000 in total losses and highlighting how vulnerable AI agents themselves are from a security standpoint.

Confirmed

• METR published an official security update disclosing two incidents this year: in March, an attacker stole a public model inference API key and burned through large amounts of credits; in May, an attacker systematically probed METR's public infrastructure, including an attempt (unsuccessful) to access internal data via an accidentally exposed endpoint.
• Details of the agent-related incident: the Dashboard code had a bug where authentication failed open and silently broke, disabling Google authentication; the attacker discovered the service via a certificate transparency list, then used prompt injection to directly induce the agent to hand over the provider's API keys, exposing API credentials worth about $600,000.
• The attacker also planted an SSH persistence backdoor, abused the keys for about three weeks, and burned roughly $600,000 in credits.
• METR stated that, following investigation, there is currently no evidence that sensitive information was leaked.

Why it matters

• The incident demonstrates the real-world risk of "AI agents themselves becoming the attack surface": fail-open authentication combined with prompt injection can make an agent voluntarily hand over high-value credentials.
• Several reposters (e.g., Gary Marcus, nptacek) connected the incident to security risks in vibe-coded apps, arguing that rapidly built, unaudited agent applications easily leave behind such vulnerabilities.
• The scale of the loss (about $600,000) and the attack duration (about three weeks before detection) show that agent systems need stronger anomaly detection and fail-closed design.

2026-09-01 ~ 2026-09-02 · 5 related posts

• METR discloses two security incidents from this year — SucceededMind · 2026-09-01
• [source] METR discloses security incidents, including a vibe-coded app with silent auth bypass — nptacek · 2026-09-01
• [source] METR loses $600k in API credits via Agent vulnerability — rohanpaul_ai · 2026-09-02
• [source] Stolen METR API key burned ~$600K in credits via fail-open agent dashboard bug — GaryMarcus · 2026-09-02

1 near-duplicate retellings: rohanpaul_ai