Autonomous Agent Liability and Forensics Matrix
How do technical failure modes like prompt injection relate to legal liability and unauthorized access under computer crime statutes?
This interactive lab presents an architectural demonstration of how autonomous agent failure modes—such as indirect prompt injection and unvalidated tool invocation—map to comparative fault and statutory liability frameworks across developers, deployers, and supervisors. The underlying mechanics explore how adversarial inputs alter execution graphs and trigger privileged API commands, while examining legal standards under statutes such as the Computer Fraud and Abuse Act (CFAA) for unauthorized or exceeded access.
The percentage distributions shown on this page (such as 45% Deployer, 20% Model Provider) are preset illustrative figures from the page markup, not empirical legal calculations or judicial determinations. In actual legal proceedings, civil or criminal liability for autonomous software incidents depends on contractual terms, indemnification agreements, intent, and factual proof under governing statutory provisions such as 18 U.S.C. § 1030.
Try a worked example
In the Failure Vector selector under Incident Causality & Attribution Lab, select 'Corrupted Tool API / Man-in-the-Middle'. The drop-down menu switches from its default 'prompt_injection' setting to 'poisoned_tool', reflecting an adversarial supply-chain or runtime tool compromise rather than a direct prompt exploit.
Statutory Frameworks and Computer Authorization
Under 18 U.S.C. § 1030(a)(2), liability arises when a party intentionally accesses a computer without authorization or exceeds authorized access, thereby obtaining protected information. The statute explicitly defines 'exceeds authorized access' under § 1030(e)(6) as accessing a computer with authorization and using that access to obtain or alter information that the accessor is not entitled to obtain or alter. 18 U.S. Code § 1030 - Fraud and related activity in connection with computers
Civil claims under 18 U.S.C. § 1030(g) require establishing damage or loss resulting from a violation, and the subsection explicitly restricts civil actions from being brought for the negligent design or manufacture of computer hardware, software, or firmware. 18 U.S. Code § 1030 - Fraud and related activity in connection with computers
Adversarial Machine Learning and Agent Security
NIST AI 100-2e2025 notes that underlying many security vulnerabilities in language model applications is the fact that data and instructions are not provided in separate channels to the model, which allows adversaries to exploit data channels to inject malicious instructions during inference. NIST AI 100-2e2025: Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
When generative models are deployed as autonomous agents equipped to invoke external tools or APIs, prompt injection attacks can hijack iterative planning loops, inducing unauthorized actions or data exfiltration from connected corporate environments. NIST CSRC: Adversarial Machine Learning Taxonomy Publication Record NIST AI 100-2e2025: Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
Sources and further reading