AI incident response

What it is for

When an AI system is attacked, the usual first questions are hard to answer: what did the model see, what did it decide, which tools did it call, with whose authority, and what did it change? G2 makes sure you can answer them, and that your team has practised.

What we look at

• Incidents specific to AI systems: prompt injection that leads to data loss or unwanted actions, a poisoned index or memory, a compromised model, tool or MCP server, and runaway cost.
• What your logs capture today, and what they need to capture for an investigation.
• Who decides to switch off an agent, and how fast that can happen.
• Your reporting duties. Under the Cyber Resilience Act, manufacturers of products with digital elements must report actively exploited vulnerabilities and severe incidents from 11 September 2026, with an early warning within 24 hours and a notification within 72 hours.^⁠[1]

Frameworks we use

CoSAI’s AI Incident Response Framework V1.0 gives AI-specific playbooks for threats such as data poisoning, model theft, prompt injection and inference attacks.^⁠[2] The OWASP GenAI Incident Response Guide 1.0 is written for security practitioners without deep GenAI expertise.^⁠[3]

What you receive

Playbooks fitted to your systems, a logging specification your engineers can implement, and a tabletop exercise that tests both with your own people.