ACSC (28 Sep): Protect your organisations’ AI services — API keys, tokens, sessions and third-party access
ASD’s ACSC alert Protect your organisations’ AI services (first published 28 September 2026) warns that malicious actors obtain unauthorised access to organisations’ AI services via compromised API keys, stolen authentication tokens, compromised user sessions, vulnerable applications, and third-party access arrangements. Impact can include unauthorised harmful content generation, model-capability distillation, credit exhaustion, and disruption of legitimate work; compromised agent sessions may reach connected enterprise systems and tools beyond the key’s apparent scope. Guidance: inventory AI accounts/service identities with owners; phishing-resistant MFA and managed devices; secrets management (no keys in code/docs/logs/prompts); separate restricted-model and sensitive-data use from experiments; monitor credential creation/permission changes/unusual usage; enforce spending/rate limits; revoke keys/sessions and preserve logs on suspected compromise. Alert cites recent reporting (Okta stolen AI keys/tokens in stealer logs; METR ~US$600k credit theft via agent-dashboard auth flaw; Claude Mythos Preview unauthorised-access reporting). Primary: ACSC alert page. Audience includes individuals/families, large organisations/infrastructure, and government.
- Product
- Organisational AI service access (API keys, auth tokens, user sessions, agent dashboards, third-party delegated access)
- Versions
- n/a (advisory; not a product CVE)
- Exploited in Australia?
- unknown
- Patch to
- Inventory AI credentials with owners; phishing-resistant MFA; secrets managers; least privilege; separate restricted models/data; monitor usage and permission changes; enforce spend/rate limits; revoke compromised keys/sessions; report via 1300 CYBER1 / ASD channels when impact suspected
Primary: ASD’s ACSC — Protect your organisations’ AI services (28 Sep 2026) · Vendor: ASD’s ACSC (alerts and advisories)