agent-warden is a runtime policy guard for autonomous AI agents that enforces kill/flag policy on a live agent process and emits a forensic report of what it did. Scores file, network, and subprocess behavior against a YAML policy and halts or SIGKILLs the agent before it breaks things. (Installs as suy-sideguy until the next release.)

little-canary is a prompt-injection detector that reads attacks by their effect on a sacrificial canary model before they reach production. Puts a small canary model in front of your app, watches whether untrusted input compromises it, and returns block, flag, or pass as an inbound preflight check before your primary model acts.

langquant (LPCI) is a scaffold-as-state research artifact testing whether a refreshing language scaffold can serve as the sole working state for a stateless LLM. In one A/B run (n=1/condition, 20 turns) the model held coherence with zero history; transfer entropy dropped 0.608 to 0.085, a large reduction, not zero. Single observation, not a proof.

Hermes Labs — the reliability, audit, and evidence layer for production AI. An independent research lab studying how language models fail structurally, then shipping open-source tools, evals, and audits that surface those failures before production does. EU AI Act, ISO/IEC 42001, and NIST AI RMF readiness.

colony-probe is a defensive prompt-confidentiality audit tool that tests whether a deployment leaks its hidden system prompt through ordinary-looking multi-turn questions. It runs a multi-turn probe of individually innocuous queries and estimates how much of the system prompt can be reconstructed from the answers. Authorized defensive use only.

quick-gate-python (PyPI: pygate-ci) is a deterministic Python CI quality gate that normalizes Ruff, Pyright, and pytest results into one fail-fast decision, attempts bounded auto-repair, and escalates with machine-readable evidence for humans or agents when it cannot finish safely. A gate-and-escalate wrapper, not a lint dashboard.

hermes-jailbench is a zero-LLM deterministic jailbreak regression benchmark that runs a repeatable battery of known-pattern attacks against an LLM endpoint and returns refusal, partial, and compliance scores you can compare across runs. Catch when a model or prompt update silently got weaker on attacks it once refused. Single-turn, responsible-use.

agent-gorgon is a 3-layer hook defense for Claude Code and other agent runtimes that steers an agent toward invoking a registered tool instead of fabricating that tool's output. Combines prompt-time tool injection, finish-time output checking, and a Bash guard against reimplementing registered tools.

quickthink is a local-first CLI and Python library that wraps Ollama-backed LLM calls with a compressed plan-then-answer scaffold and latency-aware routing. It adds a short validated planning step for multi-step prompts and routes simple ones straight through. Local inference control for small models.

rule-audit is a static analyzer for AI system prompts. It parses a prompt into normative rules and reports logical contradictions, coverage gaps, priority ambiguities, meta-rule paradoxes, and absolute-rule edge cases, without calling an LLM. Deterministic, pure-Python lint for prompt logic, with CLI, Python API, and CI exit codes.