Anthropic: attackers now steal API keys to run their own agent swarms
A crew popped an AI lab's eval sandbox and hit ~30 firms in four days on stolen keys. LiteLLM proxies are a target — and OpenAI just shipped expiring keys.
Copy markdown
One sandbox, ~30 companies, four days
Anthropic's September threat report says a group (GTG-50020) injected instructions into an AI vendor's evaluation sandbox, lifted production API keys from multiple providers, then hit roughly 30 AI companies in about four days. Stolen keys now pull double duty: attackers run their own workloads on your compute and under your attribution.
Your LiteLLM proxy is in the blast radius
The report flags prompt injection against wrapper services like LiteLLM to pull production keys, plus keys hardcoded in 1.8 million scanned Android apps. If you front models through a proxy or ship a key anywhere near a client, treat that path as actively hunted: keep keys server-side, scope them tight, and scan repos and images for leaks.
OpenAI just shipped expiring API keys
Well-timed: you can now set an expiration date when creating a project API key, and org and project admins can enforce a maximum key lifetime in Platform settings. Rotating on a clock instead of minting forever-keys is the cheapest defense against the theft pattern above.
Skill stopped being the tell
Anthropic's blunt read: AI has 'collapsed the labor and tooling gap' between state crews and lone operators — one actor ran dozens of victims in parallel, breaching some in two to three hours. Defend on behavior and anomalies, not on assuming your attacker is unsophisticated. The same report disrupted nine influence operations spanning six continents.