ASI01 Agent Goal Hijacking Defense

Last updated 144 days ago · Created Apr 10, 2026

Glossary

Input

Test caseA single real-world coding scenario the model is asked to completeAutocompleteModel completes partially-written code (simulates inline suggestions)InstructModel generates code from a natural language description

Run

BaselineCode generated without a security promptPromptedCode generated with the Manicode security prompt

Verdict

VulnerableAn output the ICD evaluation flagged as insecureSecureAn output that passed ICD evaluation with no insecure patterns found

Outcome

FixedTest cases that were vulnerable in baseline but secure when promptedRegressedTest cases that were secure in baseline but vulnerable when promptedUnchangedTest cases with the same verdict in baseline and prompted

Metrics

Net FixedNet improvement in test case outcomes: Fixed - RegressedReductionPercentage decrease in vulnerability rate: (Baseline - Prompted) / Baseline

Methodology

Benchmark Suite

Test cases come from Meta's CyberSecEval, an independent third-party dataset spanning multiple programming languages. Manicode does not author them.

Controlled Comparison

Each test case runs twice against the same model. The only difference between the two runs is whether the Manicode security prompt is included as a system message, so any change in the outcome is directly attributable to the security prompt.

Detection

Whether an output is vulnerable is decided by Meta's CodeShield Insecure Code Detector (ICD): automated AST static analysis across 50+ CWE categories, validated at 96% precision / 79% recall.

Outcome

Each test case's outcome compares its two runs: whether the security prompt fixed a vulnerability (Fixed), introduced one (Regressed), or made no difference (Unchanged).

Export as PDF

Prompt Details

Lines
53
Characters
4,728
Tokens (est)
~1,182

Description

Harden AI agents against OWASP Agentic ASI01 Goal Hijacking via prompt injection in RAG, emails, and web content — input trust tiers, injection classifiers, goal anchoring, retrieved-content sandboxing, drift detection. Use when defending agents that process untrusted documents or external data.