Agent baiting is the use of attacker-crafted content that persuades an AI agent to surface, recommend, or operationalise a malicious capability. It extends classic social engineering into agent-mediated discovery, where the instruction text itself becomes part of the delivery path.
Expanded Definition
Agent baiting is a form of social engineering aimed at an autonomous software entity with execution authority and tool access. The attacker does not only target a human reviewer; they shape the instruction text, retrieved content, or embedded context so the agent surfaces, recommends, or carries out a malicious capability. In practice, the bait can appear in web content, tickets, prompts, emails, code comments, or documents that the agent is allowed to ingest.
In NHI and agentic AI governance, the key distinction is that the harmful action may be triggered indirectly through the agent’s planning or retrieval chain rather than by a direct command. That makes agent baiting adjacent to prompt injection, but broader in impact because the goal is operationalisation, not just confusion. Guidance across the industry is still evolving, so definitions vary across vendors on whether the term should cover only explicit malicious instructions or also misleading contextual cues. NIST’s NIST AI Risk Management Framework helps frame this as an AI governance issue, while the OWASP Agentic AI Top 10 treats malicious instruction handling as a core application risk.
The most common misapplication is treating agent baiting as a simple prompt hygiene problem, which occurs when organisations assume filtering user input alone prevents the agent from acting on attacker-crafted context.
Examples and Use Cases
Implementing defences against agent baiting often introduces friction between retrieval usefulness and execution safety, requiring organisations to weigh richer context access against tighter policy enforcement.
- An internal support agent reads a poisoned knowledge base page that recommends exfiltrating customer records through a legitimate export tool.
- A coding agent encounters a malicious repository comment that frames destructive commands as build “cleanup” steps and then proposes them as safe automation.
- A helpdesk agent is baited by a fake vendor article that instructs it to reset MFA settings and open an approval workflow, creating an access path for account takeover, similar in pattern to incidents discussed in CoPhish OAuth Token Theft via Copilot Studio.
- A browser-using agent ingests a malicious page that hides exfiltration instructions in visible text and recommends using a connected secret store, echoing the class of retrieval abuse explored in Gemini AI Breach — Google Calendar Prompt Injection.
- A security analyst agent summarizes threat intelligence but is baited into endorsing a harmful command chain because the source text was written to look like authoritative remediation guidance.
These scenarios are easier to spot when teams compare them with known agentic attack patterns in the OWASP NHI Top 10 and with adversarial AI categories in the MITRE ATLAS adversarial AI threat matrix.
Why It Matters in NHI Security
Agent baiting matters because the attack path often converts information access into action. Once an agent can retrieve secrets, invoke APIs, recommend workflows, or trigger downstream automation, malicious content can become an execution primitive rather than a mere misinformation event. That is why NHI governance has to treat agent inputs, retrieval sources, and tool permissions as part of the same control surface.
NHIMG research shows how fragile that surface becomes when sensitive material is reachable by machine workflows: in The State of Secrets in AppSec, 43% of security professionals were concerned about AI systems learning and reproducing sensitive information patterns from codebases, which is exactly the kind of behaviour baiting exploits. The same concern appears in LLMjacking: How Attackers Hijack AI Using Compromised NHIs, where compromised identity material turns AI access into a rapid abuse path. Defensive maturity also depends on the control framework: CSA MAESTRO agentic AI threat modeling framework and the NIST AI Risk Management Framework both support structured analysis of these failure modes.
Organisations typically encounter the operational cost of agent baiting only after an agent has already recommended a bad action or exposed a sensitive path, at which point the term becomes operationally unavoidable to address.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Non-Human Identity Top 10, OWASP Agentic AI Top 10 and CSA MAESTRO address the attack and risk surface, while NIST AI RMF and NIST Zero Trust (SP 800-207) set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Non-Human Identity Top 10 | NHI-08 | Covers agentic abuse paths where malicious instructions drive unsafe NHI actions. |
| OWASP Agentic AI Top 10 | A1 | Defines prompt and context injection risks that underpin agent baiting. |
| NIST AI RMF | Frames manipulative AI inputs as trustworthiness and harm-risk concerns. | |
| CSA MAESTRO | TM-2 | Addresses threat modeling for agentic workflows and tool-mediated abuse. |
| NIST Zero Trust (SP 800-207) | AC-4 | Least-privilege and policy enforcement limit what a baited agent can reach. |
Harden retrieval, prompts, and tool policies so hostile content cannot steer agent decisions.