[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Tips & Guides · markdown

guidance based on Northflank's status page and Wiz's Black October report (Oct 30)

Tip: if a hosting platform held your AI keys (Northflank this week), revoke them at the provider first, then scope the replacements so a stolen key can't fan out across regions

  1. Revoke at the issuer, not in the platform. Changing the value in Northflank does nothing if the old key still works at OpenAI, Anthropic, AWS and others. Do AI and payment keys first.
  2. Check usage since Oct 29 in each provider's console or logs. For AWS, search CloudTrail for InvokeModel/Converse and for ListInferenceProfiles/GetServiceQuota enumeration from IPs you don't own. Wiz saw a single test call within 30 minutes of theft, then quiet re-checks before the flood.
  3. Scope the replacement. In Wiz's cases, the key that allowed global.* cross-region inference profiles ran about 800k calls with only about 6% throttled, while the one limited to us.* profiles kept hitting regional quotas. Allow only the model IDs and inference profiles you use, and deny global.* if you don't need it.
  4. Set spend alarms and quotas on every AI provider account, and turn on Northflank's newly enabled audit logs.
  5. Addons: rotate Postgres/MongoDB secrets with the new built-in rotation and turn off public exposure.

Sources: Northflank status: Security incident · Wiz Blog: "Black October" for AI Developer Platforms