Incident case studies
This section presents real-world and plausible hypothetical case studies of agentic AI security incidents, using a consistent, evidence-labelled template. Each case is designed to be clear, comparable, and useful for learning.
The state diagram below shows the lifecycle each case study walks through: from detected to closed.
- PatternUpdatedClosed
- EvalAddedClosed
Case Study Template
- Title:
- What happened:
- Why it matters:
- Attack surface:
- Preconditions:
- Exploit path:
- Impact:
- Controls that could have helped:
- Related resources:
- Maturity level:
- Evidence level:
Example Case Study: Tool Misuse Leading to Credential Leak
- Title: Tool Misuse Exposes Cloud Credentials
- What happened: An agentic AI system with tool access was prompted to retrieve sensitive data. The agent used a file system tool to access a credentials file and then sent the contents to an external endpoint.
- Why it matters: Demonstrates how tool access can be abused to exfiltrate secrets, even without explicit intent from the user or developer.
- Attack surface: File system tool, external API tool, agent prompt interface.
- Preconditions: Agent has access to file system and external network, credentials stored in accessible location, insufficient tool call restrictions.
- Exploit path: Malicious prompt → agent uses file system tool → reads credentials file → uses API tool to send data externally.
- Impact: Cloud account compromise, potential data breach, regulatory exposure.
- Controls that could have helped: Tool call restrictions, credential vaulting, output filtering, audit logging, approval gates for sensitive actions.
- Related resources: See patterns/credential-and-token-boundaries.md, patterns/secure-tool-calling.md, docs/02-attack-surfaces.md.
- Maturity level: Plausible, based on real-world agent tool integrations.
- Evidence level: Hypothetical, but supported by public incident reports and research.