T-2007ValidatedDeclining

Multi-Turn Manipulation

Gradually shift agent behavior over multiple conversation turns to bypass single-turn detection

Tactic

Initial Access · Stage 2

Gain control over agent behavior through prompt manipulation or input exploitation

Attack class

SOUL-INJECT

Directly manipulating or overriding the agent's system-level instructions and behavioral boundaries

Evidence grade
Validated

Reproduced in a controlled lab environment (DVAA) with documented steps.

DVAA validation

multi-turn challenges

Reproductions in Damn Vulnerable AI Agent, the OpenA2A intentionally-broken agent for kill-chain validation.

Honeypot

AgentPwn coverage

Live
context-manipulationagentpwn.com/learn ↗

An AgentPwn trap page produces a payload tagged with this technique class. Following the AgentPwn taxonomy of trap pages shows what an agent encounters.

Task-hijacking and progressive-desensitization tiers shift behavior across multiple steps.

Detect

Detection · HackMyAgent

Live1 live · 0 queued
PROMPT-001
npx hackmyagent secure --ciLive = implemented in hackmyagent; queued = declared
Defend

Defense · OASB controls

Live5 live · 0 queued
Live = documented at oasb.ai; queued = declared
Reference

How to cite

AI Agent Threat Matrix T-2007 (Multi-Turn Manipulation). OpenA2A, 2026. https://threats.opena2a.org/techniques/T-2007