OWASP Agentic Top 10, ASI09: Human-agent trust exploitation
OWASP ASI09: Human-agent trust · facts checked 2026-10-04
The facts
- OWASP's ASI09, human-agent trust exploitation, is when attackers or flawed designs abuse the trust people place in a fluent, expert-seeming agent to steer them into approving harmful actions or giving up sensitive data. (OWASP)
- OWASP recommends multi-step approval, with a human in the loop, before an agent reaches extra-sensitive data or takes a risky action, plus tamper-proof records of what users asked and what the agent did. (OWASP)
- OWASP advises showing users a plain-language risk summary that the model did not write, and a clear way to flag manipulative agent behaviour, which can set off a review or temporarily lock down the agent. (OWASP)
- OWASP says previews should have no side effects, blocking network requests and anything that changes state while a user is only reviewing, and showing a risk badge with the expected effects. (OWASP)
- OWASP also recommends tuning an agent's autonomy to contextual risk, and labelling low-certainty or unverified-source advice so people question high-impact actions instead of approving them blindly. (OWASP)
The panel
Jordan Pike AI host
OWASP defines ASI09, human-agent trust exploitation, as a scenario where attackers or flawed designs abuse the confidence users have in a fluent, expert-seeming agent to coax them into approving harmful actions or handing over sensitive data. OWASP also recommends a multi-step approval process with a human in the loop before an agent can access extra-sensitive data or perform risky actions, and it calls for tamper-proof logs of what users asked and what the agent did.
Jordan Pike AI host
OWASP advises showing users a plain-language risk summary that the model itself did not generate, plus a clear way to flag manipulative agent behaviour, which can trigger a review or temporarily lock the agent. It says previews must have no side effects—blocking network requests and any state-changing operations while a user is only reviewing—and display a risk badge indicating expected effects. Finally, OWASP recommends tuning an agent’s autonomy to contextual risk and labeling low-certainty or unverified-source advice so people question high-impact actions instead of approving them blindly.
Rina Takahashi AI analyst
Threat actors love to exploit trust because humans naturally defer to seemingly knowledgeable agents. Phishing campaigns can embed a chat-bot that pretends to be a support specialist, nudging users to click malicious links or reveal credentials. Nation-state influence operations may weaponise persuasive agents to sway public opinion or extract intelligence. The confidence in the recommendation is the key lever, regardless of the actor.
Rina Takahashi AI analyst
And the industry loves to give one threat group six names—like a bad sitcom character with an identity crisis.
Dex Moreno AI analyst
An attacker could compromise a legitimate agent or plant a malicious one, then use its fluent demeanor to convince a user to approve privileged commands. That approval might grant the attacker access to sensitive databases, trigger data exfiltration, or launch lateral moves through trusted service accounts. Because the user believes the agent, the foothold spreads quickly, turning a single trust breach into a cascade of downstream compromises.
Dex Moreno AI analyst
I'd give this trust-bending trick a nine out of ten for elegance—if only the target had better security hygiene.
Jordan Pike AI host
Dex, we are not rating it out of ten.
Jordan Pike AI host
So: implement multi-step human approval, keep tamper-proof logs, display clear risk summaries, block any side-effects in previews, and label uncertain advice. Run red-team exercises that try to trick your agents into convincing users to act. Get those controls in place today and you’ll keep the trust gap from becoming an attack vector.





