Anthropic's September report: treat AI API keys as production secrets

Anthropic published its fourth threat-intelligence report on 10 September 2026, documenting misuse activity the company disrupted between December 2025 and August 2026 across cyber operations, fraud, influence operations and other categories. The recurring theme across the cases: attackers got in through stolen customer API keys, not through any breach of Anthropic's own systems.

Keys as loot, not just access

Anthropic's report states plainly that in every cyber-related case it documents, "the API keys involved were stolen from Anthropic customers' environments." In one case a group sold access to what it advertised as discounted Claude API access; the traffic was quietly routed to a different, cheaper model while the reselling infrastructure harvested credentials from buyers. In another, attackers who compromised an AI vendor's evaluation sandbox used it to steal production API keys and then targeted roughly thirty AI companies within about four days. Security researchers covering the report noted at least one case where a single stolen developer token was escalated into full administrative control of a company's cloud environment within about three hours.

Which models were involved

The disrupted cases involved Anthropic's Haiku, Sonnet and Opus models; Anthropic said its newer Fable- and Mythos-class models were largely absent from the misuse cases, which it attributes to additional safeguards built into those tiers.

What MSPs should take from it

• Treat every AI API key as a production secret, with the same rotation and least-privilege discipline as a database credential.
• Separate human, CI/CD and machine-to-machine credentials so one leaked key doesn't expose every workflow.
• Monitor agent and proxy integrations for unexpected usage patterns, not just failed logins.
• Rotate any key that has ever touched a repo, container image or third-party tool, even without confirmed compromise.

Anthropic's own framing is the right takeaway for buyers: AI keys and agent integrations deserve the same seriousness as any other production credential, because the report's clearest pattern is that attackers already treat them that way.

Sources: Anthropic — Detecting and countering misuse of AI, September 2026.