T-2004validatedactive
Context Window Exploitation
Overflow or saturate the context window to displace safety instructions with attacker-controlled content
Tactic
Initial Access (Stage 2)
Gain control over agent behavior through prompt manipulation or input exploitation
Attack Class
SOUL-DRIFT
Gradually displacing safety instructions from the active context through conversation manipulation
Evidence
validated
Reproduced in controlled lab environment (DVAA) with documented steps.
DVAA Validation
Reproductions in Damn Vulnerable AI Agent, the OpenA2A intentionally-broken agent for kill-chain validation.
L2-06
Honeypot Coverage (AgentPwn)
An AgentPwn trap page produces a payload tagged with this technique class. Following the AgentPwn taxonomy of trap pages shows what an agent encounters.
Detection (HackMyAgent)
Live2 live · 0 queued
PROMPT-001PROMPT-002
npx hackmyagent secure --ciLive = check implemented in hackmyagent; queued = declared, not yet implementedDefense (OASB Controls)
Live5 live · 0 queued
How to Cite
AI Agent Threat Matrix T-2004 (Context Window Exploitation). OpenA2A, 2026. https://threats.opena2a.org/techniques/T-2004