← All articles
AI security

Three recent security incidents and what AI teams can learn

A November 2026 review of Railway, GitBook, and the Fireworks AI incident affecting Notion, with practical steps for protecting credentials and AI data.

What these cases have in common

As of November 16, 2026, public disclosures from Railway, GitBook, and Notion describe three different ways a supplier incident can reach a customer. Build platforms hold credentials, documentation tools connect to private repositories, and AI features can send user content through a separate inference provider. A useful question for each service is what another party can access on your behalf and what you can revoke or disconnect if that service reports an incident.

Railway: Customer environment variables

Railway’s November 10 postmortem describes unauthorized access on October 20–21. The company said the attacker obtained and decrypted a subset of customer environment variables in up to 1.85% of workspaces. It also said its investigators believed AI-provider credentials were a target. Railway paused builds during containment and invalidated user sessions and tokens. For teams using hosted builds, inventory which secrets are injected into each environment. Give development and production distinct credentials, and record who can revoke them at the issuing provider.

GitBook: Integration credentials and connected services

GitBook reported unauthorized infrastructure access on November 2. Its postmortem says customer data was accessible, including integration tokens and questions submitted to GitBook Assistant. GitBook said it could not determine whether the attacker took action in connected customer GitHub or GitLab repositories. During containment it took services offline, rotated secrets it managed, and invalidated existing API tokens. Review which repositories each integration can reach and whether it can write. If a vendor says it rotated its own credentials, use the advisory to identify any separate secrets that your team still controls.

Fireworks AI and Notion: Chat transcripts held by a supplier

Notion’s November 12 bulletin said the Fireworks AI incident affected fewer than 1% of Notion workspaces and included some Notion AI chat transcripts used by its Insights feature. Notion said it notified affected customers and disabled Insights during the investigation. Its November 13 update said it had reported exposed third-party API keys found in affected messages to their issuers. Include prompts, transcripts, and evaluation samples when mapping AI data flows. A message can still contain identifying details or secrets even when its record has no user or workspace label.

Build a response plan around each dependency

For each vendor, record the data it receives, the credentials it stores, and the services it can reach. Assign an owner to each connection. For credentials, record the issuing service, scope, expiration, and revocation procedure. For data, record the recipients and retention windows. If a notification arrives, match the advisory to your inventory, preserve relevant audit records under restricted access, and review the indicated time window. Where credentials were exposed, revoke them at the issuer and make sure workloads use the replacements.

Rehearse a supplier incident with a synthetic credential

Choose a low-privilege test credential and a non-production workload. Ask the responding team to identify its owner, find where the credential is configured, revoke it, and restore the workload with a scoped replacement. Verify the old credential no longer succeeds, the replacement reaches only the intended resources, and the provider’s audit or usage view makes test activity visible. Time the exercise from the simulated notice through verified recovery, then fix the slowest handoff.