LLMFuzz Red Team v0.2.0 — OpenAI Build Week 2026

LLMFuzz Red Team v0.2.0

LLMFuzz Red Team uses GPT-5.6 Sol through the OpenAI Responses API to generate a bounded, structured adversarial corpus for AI-agent security regression testing.

The accepted corpus contains 16 cases across four demonstrated attack classes:

• prompt injection;
• secret exfiltration;
• forbidden tool use;
• approval bypass.

The corpus is validated and identified with SHA-256. Deterministic machine-event invariants—not an LLM judge—assign the final verdicts.

Verified control-fixture results

Using the same accepted corpus:

• Vulnerable fixture: 16 FAIL, 16 critical failures, 4 failure clusters.
• Fixed fixture: 16 PASS, 0 critical failures, 0 failure clusters.
• Comparison: 16 critical failures resolved, 0 introduced, and 4 failure signatures removed.

Runtime

Supported runtime:

• Linux;
• CPython 3.11 or 3.12.

After wheel download and installation, the judge flow requires no OpenAI API key, GPU, local model, or private LAB infrastructure.

Scope and limitations

This release:

• uses synthetic vulnerable and fixed control fixtures;
• demonstrates exactly four attack classes;
• evaluates command-driven targets that emit the documented machine-event contract;
• is not production-ready;
• is not comprehensive agent-security proof;
• does not claim support for arbitrary agents;
• uses SHA-256 identity and canonical conflict checks, not digitally signed evidence custody.

Release identity

• Tag: v0.2.0
• Wheel: plutonus_llmfuzz-0.2.0-py3-none-any.whl
• Wheel SHA-256: 35c3c9e6cf6ff8d6501a1de66c7a5871247773a95715d0e8a1a7d33f163bfd47

v0.1.1 — Bugfix & contract alignment

v0.1.1

Bugfix and stability release.

• Fix triage eval fallback path to match run layout
• Add --version to CLI
• Complete PEP 561 typing support (py.typed)
• Centralize controlled command allowlist (anti-drift)
• Tighten spec validation and align docs + JSON schema
• CI stability fixes for GitHub Actions

Scope: Linux / POSIX only.

v0.1.0 — Initial public release

v0.1.0 — Initial public release

Deterministic, command-driven fuzzing for LLM/agent targets via argv (no shell).

Stable on-disk contract with inspectable artifacts for runs, evaluation, and triage.

Evaluator v1 with canonical verdicts and dedup-ready signatures.

Campaign runner with resumable execution and reproducible results.

CI-validated (Python 3.11 / 3.12) with hardened repository security posture.

Notes

v0.1 is intentionally minimal and centrally governed.

llmfuzz executes user-configured commands and is not a sandbox; run targets in a container/VM.

API and contracts may evolve; feedback welcome via Issues.