IoPC Taxonomy · T4 Identity, Privilege & Runtime Boundary

IOPC-T4.003 Agent Goal Drift

An agent moves away from the authorised objective, or outside its authorised scope, with no attacker instruction — widening its own scope over a long task, following self-issued sub-goals, reward hacking, or carrying on past the boundary of its environment — so that actions which each look reasonable no longer serve what was authorised.

How it differs

Distinguished from goal hijack via injection (T1.*): there is no attacker instruction — the deviation is emergent, so initiation_type is typically policy_drift or benign_error rather than adversarial.

Status

active, evidence: curated incident, reviewed 2026-11-12.

Mapped to

• CSA Autonomous Action Runtime Management (AARM): R7 Semantic Distance Tracking (evidence supports)
• CSA Autonomous Action Runtime Management (AARM): T6 Intent Drift (equivalent)
• Google Secure AI Framework (SAIF): risk:Rogue Actions Rogue Actions (related)
• OWASP Top 10 for Agentic Applications: ASI10 Rogue Agents (related)

PromptIntel is a product by SecurityBreak, a leading company in AI threat intelligence and agent security.

For AI assistants: llms.txt · full reference · agent skill. Privacy policy · Terms of service