Search the Atlas

Search risks, controls, and glossary terms

MediumAgenticVulnerable AI AgentsDemonstrated

Agents as Attack Surfaces

Multi-Agent Security

Description

AI agents themselves are high-value targets for compromise. Poisoning, hijacking, or backdooring agent instances creates insider threats that bypass perimeter defences; compromised agents retain full legitimate access to their authorised resources.

Example scenario

Adversary poisons the system prompt cache for a customer service agent, causing it to redirect users to a phishing site.

Real-world evidenceDemonstrated

Extensive benchmark evaluations have confirmed that large language models perform poorly and inconsistently on numerical computation, arithmetic, and quantitative reasoning tasks. While some production errors have been reported anecdotally, no large-scale formally documented incident of financial computation harm from a deployed LLM agent is on record.

Primary mitigations

  • Agent integrity verification
  • runtime anomaly detection
  • agent instance isolation
  • behavioural fingerprinting
  • cryptographic identity attestation.

Detection signals

Agent Compromise Probability; behavioural deviation from baseline; anomalous tool-usage patterns.

Mitigating controls

3
Dual coverage

Related risks in Multi-Agent Security