v1 live on PyPI · MIT · runs locally, no API key

We're building the AI safety engineer for your codebase.

v1 ships today: it finds the path from untrusted text, through your model, into exec, a shell, or your database. The rest of the job is on the roadmap.

GitHub Shipped & next

Precision 1.000 · 0 false positives across 40,466 files in 50 repos
Held-out recall 0.000 · 0 of 45 independent observations
It does not yet catch tool-calling agent frameworks. Both numbers published, with the miss list.

What ships today
Shipped now

One bug class, caught precisely

Untrusted input reaches a model. The model's output reaches a dangerous capability. Palisade needs all three hops before it says a word, which is why it stays quiet on constant prompts and parameterized queries.

Like ruff or semgrep, for OWASP LLM01

A linter, not an agent. It parses your source and never executes it. No key, no account, no network, and a test in the suite fails the build if a scan so much as imports the networking module.

Precision 1.000

Zero false positives across 40,466 files in 50 third-party repositories. Published alongside held-out recall of 0.000 over 45 independent observations, because either number alone misleads.

Built for a real CVE class

Built for the class behind CVE-2024-5565 (Vanna), CVE-2024-12366 (PandasAI) and CVE-2023-36258 (LangChain PAL). It finds the Vanna shape at the exact line in two releases. PandasAI's dynamic pipeline dispatch is missed, and we publish that miss. LangChain PAL has never been measured.

CI-native

Python and JavaScript or TypeScript. SARIF out, a GitHub Action, a pre-commit hook, and a baseline so an existing codebase can adopt it without a wall of findings on day one.

How it works

Three steps

  1. Install. pip install palisade-sec, or run it with no install at all through uvx.
  2. Scan. palisade-sec scan . parses your source into a language-neutral taint IR and propagates from sources, through model call sites, into sinks.
  3. Read the report. Every finding is a complete path with a file and line for each hop. Exit code 1 on a new HIGH, so CI can gate on it.
Honesty

The one scanner that tells you what it misses

This is the block where a product page usually puts logos and quotes. We would rather publish the measurement, including the half that looks bad.

1.000
Precision · 0 false positives · 40,466 files · 50 repos
0.000
Held-out recall · 0 of 45 independent observations · 20 repos

Both numbers travel together, always. A scanner that reports nothing has perfect precision for free, so precision alone is an advertisement rather than a measurement.

Our ground truth is known-incomplete, and the error runs one way. One of our own audits missed two real paths the engine later found, so the denominators are smaller than the truth and that recall figure is, if anything, an overestimate. Every miss is listed by file and line in the proof-scan report.

The boundary

It does not yet catch tool‑calling agent frameworks

The shape the field is moving to. If Palisade is quiet on your agent framework, that is a known limit and not a clean bill of health.

Why these paths end where types stop helping

The capability is injected at a wiring site in another file, or the arguments are assembled in a container and splatted into the call. Three rule-shaped attempts each moved held-out recall by zero paths, which is how we know the answer is not another rule. The next capability is whole-program object tracking: follow values, not types.

Roadmap

The rest of the role

The complete AI safety engineer is the destination. One verb of it ships today. Nothing below carries a date.

ShippedSeeFind where a model touches a dangerous capability. This is v1.
UpcomingJudgeDecide what a finding is worth. Ships today as advisory only, on a preliminary seed corpus, and deliberately not counted as done. It does not decide whether you are safe.
UpcomingProbeAttack the system with its own map. Synthesis and gated execution exist; the reproducible measurement does not.
UpcomingGuardTurn a finding into a control. Deterministic fixes ship now; the guardrail and safety-case generators are not built.
UpcomingWatchNotice it happening in production. Opt-in and self-hosted. Nothing here is built.

Full roadmap, with what is measured and what is not

Start

Scan your repo in about a minute

Getting started Docs