agentsastLast reviewed 2026-09-13

Glossary

Direct answerShort, precise definitions of the terms that decide whether an AI audit result means what it appears to mean: false positive rate, precision vs recall, agentic scanning, LLM plus fuzzing, hallucinated vulnerabilities, triage burden, benchmark contamination, human-in-the-loop, prompt injection in auditing pipelines, and responsible disclosure of AI-found bugs.
False positive rate
The share of reported findings that are not real, exploitable bugs. Benchmarks that count only recall against a curated bug list cannot measure it.
Precision vs recall
Precision is valid findings divided by all findings reported; recall is known bugs found divided by all known bugs. A useful evaluation reports both.
Agentic scanning
An LLM that plans, reads files, runs tools and tests, and iterates over a codebase, instead of answering one prompt over pasted code.
LLM plus fuzzing
The model proposes harnesses, seeds or mutation patterns; the fuzzer supplies ground truth by crashing or violating an oracle.
LLM plus symbolic execution or formal verification
The model generates properties, invariants or code; a solver or prover checks them, so the model's output is accepted only when a machine confirms it.
Hallucinated vulnerabilities
Confident, plausible findings that do not reproduce. The curl project ended its bug bounty on 31 January 2026 after confirmed-report rates fell below five percent.
Triage burden
The human time needed to validate candidate findings. zkSecurity reported more than a thousand candidates across two hundred projects as its main bottleneck; Codex Security reported more than ten thousand high findings in its first month.
Benchmark contamination
The model has seen the audit report or the bug in training, so a benchmark hit measures recall of memory rather than discovery.
Human-in-the-loop
A named person validates exploitability, severity and disclosure before a finding is reported to a client or maintainer.
AI-assisted audit vs AI audit
An AI-assisted audit is a human-signed report in which AI widened coverage; an AI audit is machine output with no accountable reviewer.
Prompt injection in auditing pipelines
Repository contents (commit messages, comments, PR descriptions, issue text) can carry instructions that hijack an agent running with CI privileges.
Responsible disclosure of AI-found bugs
The same coordinated-disclosure rules as human findings, plus a duty to send only validated, reproducible reports so maintainers are not flooded.
Continuous scanning and run-count coverage
Running AI analysis repeatedly, on every commit and whenever models improve, on the premise that LLM bug finding is non-deterministic and coverage compounds with runs.
Proof-of-concept harness
Tooling that turns a candidate finding into an executable exploit or failing test, such as zkao's cheatVM for Aleo or Codex Security's sandbox.
Threat model file
A scope document the tool reads before scanning (zkao's zkao.md, Codex Security's generated threat model) stating trust boundaries, adversaries and out-of-scope behaviour.
Severity calibration
Whether the tool's assigned severities match what an expert would assign. In the CIRCL study, four of seven AI severities were too high and one critical was rated medium.