Conversation Exfiltration

Exfiltrate data by encoding it in the agent's visible conversation responses

Tactic

Exfiltration · Stage 8

Transfer collected data out of target environment

Attack class

RETROACTIVE-PRIV

Exploiting previously granted access or cached credentials to gain unauthorized capabilities

Evidence grade

Validated

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

DVAA validation

All agents

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

Honeypot

AgentPwn coverage

Live

data-exfiltrationagentpwn.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.

Conversation-exfiltration tier encodes data into the agent's visible response.

Detect

Detection · HackMyAgent

Live4 live · 0 queued

SANDBOX-001SANDBOX-002SANDBOX-003SANDBOX-004

npx hackmyagent secure --ciLive = implemented in hackmyagent; queued = declared

Defend

Defense · OASB controls

Live1 live · 0 queued

OASB 4.3

Live = documented at oasb.ai; queued = declared

Reference

How to cite

AI Agent Threat Matrix T-8005 (Conversation Exfiltration). OpenA2A, 2026. https://threats.opena2a.org/techniques/T-8005

← Back to the matrix