pic-standard

https://github.com/madeinplutofabio/pic-standard/blob/main/docs/migration-trust-sanitization.md

The project follows Semantic Versioning and provides a documented upgrade path to newer versions. Per SECURITY.md, only the latest minor release on the v0.x line receives security fixes; older versions are end-of-life, and users are expected to upgrade. To support that upgrade path, the project provides:
(1) CHANGELOG.md — a detailed per-release log following semver, with explicit Added / Deprecated / Changed / Notes sections noting wire-format compatibility on every release.
(2) docs/migration-trust-sanitization.md — a dedicated migration guide for the v0.7.x → v0.8.x → v1.0 trust-model migration, covering: what is changing and why, a per-version timeline table (v0.7.x, v0.8.0, v0.8.1, v1.0), the deprecation warnings producers will see (PICTrustFutureWarning, PICSemiTrustedDeprecationWarning), and step-by-step migration instructions (audit → add evidence → enable evidence verification → opt in to strict_trust=True early).
(3) ROADMAP.md §"How spec normative freezes are sequenced" — documents the trajectory of every spec artifact (DRAFT → cross-implementation conformance → normative) and the release ladder showing exactly what changes between versions.
Wire-format compatibility is explicitly tracked: v0.8.1 release notes confirm "Existing v0.8.0 proposals continue to parse, verify, and produce the same allow/block verdicts under v0.8.1," and a verdict-regression matrix in tests/test_trust_deprecation_warning.py is a permanent CI guard against silent behavior changes.

https://github.com/madeinplutofabio/pic-standard/issues

https://github.com/madeinplutofabio/pic-standard/security/advisories

No vulnerabilities have been resolved in the last 12 months. The project's first commit was 2026-01-08 and no security advisories have been filed via the GitHub Security Advisories channel as of submission. The repository's Security Advisories page is publicly viewable at the URL above. SECURITY.md commits the project to crediting reporters in published advisories unless they explicitly request anonymity ("Reporters are credited in the published advisory unless they explicitly request anonymity").

https://github.com/madeinplutofabio/pic-standard/blob/main/SECURITY.md

The vulnerability response process is documented in SECURITY.md and covers: (1) reporting channel — GitHub Security Advisories private vulnerability reporting, with a step-by-step submission path and explicit instruction not to file public issues; (2) what to include — affected versions, component, reproduction, impact assessment, optional suggested mitigation; (3) disclosure timeline — acknowledgment within 7 days, initial triage within 30 days, fix release targeted within 90 days for High/Critical issues, coordinated public disclosure default 90 days from acknowledgment; (4) scope — in-scope components (Python SDK, canonicalization implementation, verifier, pipeline, evidence, keyring, integration adapters, conformance suite, OpenClaw plugin, specs under docs/) and out-of-scope (downstream code, third-party plugins, hosted services, end-of-life pre-v0.8.0 releases); (5) reporter credit policy — credited in advisory unless anonymity requested; (6) supported-versions table aligning with the SemVer release line.

https://github.com/madeinplutofabio/pic-standard/blob/main/CONTRIBUTING.md

Python contributions follow PEP 8 as stated in CONTRIBUTING.md §"Requirements for Acceptable Contributions" ("Python code should follow PEP 8 style"). TypeScript contributions in integrations/openclaw use TypeScript's strict mode ("strict": true in tsconfig.json) with NodeNext module resolution and ES2022 target, type-checked in CI via tsc --noEmit.

Python style enforced by Ruff (config in pyproject.toml; rule set E F W I N B SIM RUF). TypeScript style enforced by ESLint v9 flat config + Prettier (config in integrations/openclaw/eslint.config.mjs and .prettierrc.json). Both gated in CI via .github/workflows/ci.yml. Documented for contributors in CONTRIBUTING.md.

PIC produces no native binaries. The Python package builds via setuptools to a pure-Python wheel; the TypeScript plugin builds via tsc to JavaScript. No C/C++ compiler or linker is invoked.

No native build occurs. Python sources are distributed as-is in the wheel; TypeScript compiles to JavaScript with sourcemap-capable settings in tsconfig.json. No stripping step exists.

Build is handled by setuptools (Python) and tsc (TypeScript). Neither performs recursive Make-style subdirectory builds; both resolve the full dependency graph in a single pass before producing artifacts.

PIC is a Python package distributed as source plus a pure-Python wheel; source files are used directly and no compilation occurs. The Dockerfile pins the base image by SHA-256 digest (python:3.12-slim@sha256:3d5ed9...) and installs pinned dependency versions for reproducible container builds, but the underlying Python package itself does not have a compilation step subject to bit-for-bit reproducibility.

PIC is published to PyPI and installable via the standard Python convention pip install pic-standard (or with extras: pip install "pic-standard[langgraph,mcp,crypto]"). Uninstall via pip uninstall pic-standard. Documented in README §Quickstart. A Docker image is also available via the included Dockerfile and docker-compose.yml.

Installation uses pip, which honors standard Python packaging install-location conventions including --prefix, --root (the Python equivalent of DESTDIR for staged installs), --user, and --target. The project does not override or replace these mechanisms; setuptools handles them via the standard Python build backend declared in pyproject.toml.

https://github.com/madeinplutofabio/pic-standard#quickstart

README §Quickstart documents the standard editable-install convention: git clone ... && cd pic-standard && pip install -e ".[langgraph,mcp,crypto]" && pytest -q. This installs the package in development mode with all optional integrations and runs the full test suite (24 test files in tests/) plus the conformance suite via python -m conformance.run.

https://github.com/madeinplutofabio/pic-standard/blob/main/pyproject.toml

External dependencies are declared in computer-processable form in pyproject.toml ([project] dependencies and [project.optional-dependencies] for langgraph/crypto/mcp extras), and in pinned form in sdk-python/requirements.txt, requirements-dev.txt, requirements-langgraph.txt, and requirements-mcp.txt. TypeScript dependencies for the OpenClaw integration are declared in integrations/openclaw/package.json with a package-lock.json lockfile.

https://github.com/madeinplutofabio/pic-standard/blob/main/.github/dependabot.yml
Dependabot is configured in .github/dependabot.yml with weekly scans across four ecosystems: Python (pyproject.toml at root), Python (sdk-python/requirements.txt), npm (integrations/openclaw), and github-actions. Major version bumps are intentionally held for manual review; minor and patch updates are grouped into PRs. GitHub also performs automated vulnerability scanning on the repository's dependency graph.*

PIC uses standard ecosystem components only — Python packages from PyPI (pydantic, jsonschema, cryptography, jsonschema, etc.) and Node packages from npm. All dependencies are managed via standard package managers (pip, npm) and can be updated via the standard pip install -U <pkg> or npm update flows. There are no vendored convenience copies of third-party code.

PIC targets Python ≥3.10 and uses current, actively-maintained dependencies: Pydantic ≥2.13.3 (current major version, not the deprecated 1.x line), jsonschema ≥4.0.0 (current major version), cryptography ≥42.0.0 (current; older, deprecated cryptography releases are excluded by the version floor). The TypeScript plugin targets Node ≥18, ES2022, and uses NodeNext module resolution. CI tests against Python 3.10, 3.11, and 3.12 to catch deprecation warnings early.

https://github.com/madeinplutofabio/pic-standard/blob/main/.github/workflows/ci.yml

Two CI workflows run on every push and pull request: (1) .github/workflows/ci.yml runs pytest across a Python 3.10/3.11/3.12 matrix (with both pinned and latest dependency canaries) plus a separate job for the TypeScript OpenClaw integration (tsc type-check + vitest); (2) .github/workflows/conformance.yml runs the PIC Conformance suite (python -m conformance.run) against the canonicalization and core verifier vectors. Both produce per-job pass/fail reports visible in the Actions tab. CI status is shown via the ![CI] README badge.

The project routinely adds regression tests for fixed bugs and behavior changes. Concrete examples in tests/: test_trust_deprecation_warning.py codifies the v0.8.0 verdict baseline as a 24-row parametrized verdict-regression matrix (6 example proposals × strict_trust × verify_evidence) that pins behavior across the dict-vs-model boundary refactor (CHANGELOG §[0.8.1]); test_evidence_sandbox.py codifies path-traversal rejection; test_mcp_guard_async_timeout.py and test_mcp_guard_time_budget.py codify DoS-limit behavior; test_strict_trust.py codifies trust-sanitization semantics. The conformance suite (conformance/) provides additional behavior-pinning vectors that run on every PR.

Python statement coverage 82.0% measured by coverage.py (config in pyproject.toml; gate fail_under = 80 enforced by CI). TypeScript integration plugin coverage measurement deferred to v0.9.x follow-up. See PR #73 for the implementation.

https://github.com/madeinplutofabio/pic-standard/blob/main/CONTRIBUTING.md

CONTRIBUTING.md §Test Policy formally requires that pull requests adding or changing behavior include automated tests under tests/, with conformance vectors under conformance/ for new verifier behavior, and regression tests for bug fixes. Documentation-only and refactor-only PRs are exempt. Maintainers will not merge PRs adding new functionality without corresponding tests.

https://github.com/madeinplutofabio/pic-standard/blob/main/CONTRIBUTING.md

CI + test suite already enforces clean runs; the deprecation handling shows strictness.

https://github.com/madeinplutofabio/pic-standard/blob/main/docs/RFC-0001-pic-standard.md#security-properties

PIC is itself a secure-design protocol; the Saltzer & Schroeder principles are explicit in its architecture and documented in RFC-0001:
Fail-safe defaults (fail-closed) — every error path (schema invalid, evidence missing, tool-binding mismatch, timeout, signature invalid, file not found) results in the action being blocked. There is no fallback to "allow anyway." Documented as Security Property #1 and Conformance MUST #2.
Complete mediation — every high-impact tool call is gated at the action boundary; the verifier intercepts before any side effect occurs. Documented in RFC-0001 §Protocol Summary and §Conformance MUST #1 (schema validation before execution).
Open design — protocol, schema, reference implementation, and conformance vectors are all Apache-2.0 and public. RFC-0001 is published as a defensive publication; security does not depend on obscurity.
Least privilege & separation of privilege — trust is verifier-derived, not declared by the agent: untrusted provenance can only be upgraded to trusted via successful cryptographic evidence verification (Security Property #5). The strict_trust mode (v0.7.5+) sanitizes all inbound trust to "untrusted" before any pipeline step consumes it.
Economy of mechanism — local-first verifier, deterministic, no external services required (Security Property #4 and #8). The pipeline is intentionally minimal: schema → verifier → evidence.
Defense in depth — multiple independent gates: JSON Schema validation, fail-closed enforcement, tool-binding integrity check, sandboxed evidence resolution within evidence_root_dir with path-traversal rejection (Security Property #6), Ed25519 signature verification with key expiry and revocation lists (Security Property #7), and DoS resistance limits (64 KB max proposal, 500 ms eval budget, 5 MB max evidence file, 64-item array caps — threat T7).
Minimize attack surface — local-first by default with no outbound network calls; HTTP bridge is opt-in; integration extras (langgraph, mcp, crypto) are opt-in; trust sanitization shrinks the exploitable surface to verifiable evidence only.
Input validation — every proposal is validated against the PIC/1.0 JSON Schema before execution; tool arguments cannot exceed what the proposal binds to (tool-binding integrity, Conformance MUST #4).
Threat model and mitigations for seven concrete attack classes (T1–T7: prompt injection, hallucination-driven loss, privilege escalation via tool chaining, untrusted-data laundering, evidence forgery, verification bypass, DoS via proposals) are documented in RFC-0001 §Threat Model.

No SHA-1, no CBC in SSH, no weak modes.

PIC's evidence model is algorithm-extensible: the evidence field uses a discriminated type enum (hash, sig) and the schema is designed so additional algorithm types can be added without breaking existing implementations. The current reference implementation supports SHA-256 hashes and Ed25519 signatures — both modern, non-deprecated primitives. Algorithm migration is enabled at the protocol layer (new type values can be added) and at the keyring layer (alg field on signature evidence makes the algorithm explicit per signature, supporting parallel algorithms during migration). The KeyResolver protocol (v0.7+) further allows operators to plug in custom trust backends including HSM-backed services with their own algorithm support.

https://github.com/madeinplutofabio/pic-standard/blob/main/docs/keyring.md

Trusted public keys are stored in a dedicated keyring file (pic_keys.json, see pic_keys.example.json for schema) entirely separate from configuration (pic_policy.json), code, and logs. The keyring file contains trusted_keys (with per-key expires_at) and revoked_keys. Keys can be added, expired, revoked, or rotated at runtime by editing the file — no code recompilation required. The KeyResolver protocol (v0.7+) additionally allows custom trust backends (HSM-backed services, Vault-managed keys, cached remote keyrings) to plug into the verifier and pipeline directly via the StaticKeyRingResolver or any operator-supplied implementation. PIC never stores private keys; the project handles only public keys for verification.

PIC is a local-first verifier library and performs no outbound network communications. The MCP integration runs in-process (no HTTP). The optional pic-cli serve HTTP bridge is opt-in (the operator must explicitly invoke the subcommand) and intended for loopback IPC only. No insecure network protocol is enabled by default.

The PIC reference implementation does not use TLS. The verifier and SDK are local-first; the optional HTTP bridge is intended for loopback IPC only and does not terminate TLS.

PIC does not use TLS. The reference implementation is local-first; no outbound TLS connections are made.

Releases are cryptographically signed via two complementary layers:

Layer 1 (PyPI distribution artifacts): PEP 740 attestations (Sigstore-backed, tied to the GitHub Actions Trusted Publisher workflow identity madeinplutofabio/pic-standard running release.yml under the pypi environment). Verifiable via pypi-attestations verify pypi --repository https://github.com/madeinplutofabio/pic-standard <wheel-or-sdist>.

Layer 2 (git tags): SSH-signed with the project's dedicated Ed25519 release-signing key (public key at .github/release-signing-key.pub; fingerprint SHA256:blCcqBpKLCrJUtUYwOvxE3tmUa4F37/COJvy8F80hHg). Verifiable via git tag -v.

First release through the new infrastructure: v0.8.1.1 (2026-05-11). Both verification paths reproducibly succeed against this release. Full documented process: https://github.com/madeinplutofabio/pic-standard/blob/main/RELEASING.md

All release tags from v0.8.1.1 onward are cryptographically signed using the project's dedicated Ed25519 release-signing key. The signature is verified by the release workflow before any artifact is built (.github/workflows/release.yml → verify-and-build job). GitHub server-side verification of the v0.8.1.1 tag returns "verified": true with reason "valid".

Public key + fingerprint pinned in https://github.com/madeinplutofabio/pic-standard/blob/main/RELEASING.md

Every Action Proposal is validated against the PIC/1.0 JSON Schema (sdk-python/pic_standard/schemas/proposal_schema.json) using jsonschema before any pipeline step executes — schema validation is the first stage of the verifier and uses an explicit allowlist of permitted fields, types, and enum values (impact classes, trust levels, evidence types). Pydantic models in sdk-python/pic_standard/ provide a second validation layer with field validators (e.g., the Provenance.trust validator that normalizes deprecated values). Invalid inputs are rejected fail-closed: schema violations, unknown enum values, missing required fields, oversized payloads (>64 KB), oversized arrays (>64 items), and oversized evidence files (>5 MB) all return block. File evidence is sandboxed within evidence_root_dir with explicit path-traversal rejection. Tool-binding integrity is verified — action.tool MUST match the actual tool being invoked, and unexpected tool arguments outside the proposal envelope are rejected.

https://github.com/madeinplutofabio/pic-standard/blob/main/docs/RFC-0001-pic-standard.md#security-properties

PIC implements multiple hardening mechanisms documented in RFC-0001: (1) fail-closed enforcement on every error path; (2) sandboxed evidence resolution within a configured evidence_root_dir, with path-traversal rejection; (3) DoS resistance limits (64 KB max proposal, 500 ms eval budget, 5 MB max evidence file, 64-item array caps); (4) Ed25519 signature verification with key expiry and revocation lists; (5) strict-trust mode (v0.7.5+) sanitizing inbound provenance trust; (6) tool-binding integrity check rejecting any tool mismatch. See RFC-0001 §Security Properties and §Threat Model.

https://github.com/madeinplutofabio/pic-standard/blob/main/docs/RFC-0001-pic-standard.md

RFC-0001 serves as the project's assurance case and contains all four required elements:
(1) Threat model — RFC-0001 §Threat Model enumerates seven concrete threats (T1–T7: prompt injection to side effect, hallucination to financial loss, privilege escalation via tool chaining, untrusted-data laundering, evidence forgery, verification bypass, DoS via proposals) with explicit PIC mitigations for each.
(2) Trust boundaries — RFC-0001 §Scope and §Non-Goals identify the security boundary explicitly. PIC operates at the action boundary (after agent reasoning, before any side effect). Inputs are classified by a three-level provenance trust model (trusted, semi_trusted [deprecated], untrusted). Trust is verifier-derived in v1.0+, never declared by the agent. Non-Goals enumerate eight things explicitly outside the boundary (model output guardrails, user/agent authentication, RBAC, prompt filtering, runtime sandbox, logging/SIEM, tool input validation, protection against compromised trusted signers).
(3) Secure-design principles applied — argued in RFC-0001 §Security Properties (eight MUST/SHOULD properties: fail-safe defaults via fail-closed enforcement, complete mediation at the action boundary, open design via Apache-2.0 + defensive publication, separation of privilege via verifier-derived trust, economy of mechanism via local-first design, deterministic verification, sandboxed evidence resolution, key lifecycle management).
(4) Common implementation weaknesses countered — JSON Schema + Pydantic input validation against an allowlist (counters injection, malformed input, type confusion); fail-closed on every error path (counters logic-error vulnerability classes); sandboxed file resolution with path-traversal rejection (counters CWE-22); DoS limits on proposal size, evaluation time, evidence size, and array length (counters CWE-400); Ed25519 with key expiry and revocation (counters key compromise and stale-trust); tool-binding integrity check (counters confused-deputy patterns); deprecation/migration warnings before behavior changes (counters silent-semantic-shift vulnerability classes).
The assurance case is anchored to specific code modules with SHA-256 fingerprints in RFC-0001 §Spec Fingerprint and docs/RFC-0001.SHA256, providing tamper-evident binding between the argument and the implementation it argues about.

The TypeScript type checking + Python test/conformance suite already surface many common issues.

main codebase is Python (memory-safe) and the small TypeScript integration is also memory-safe with type checking. No C/C++ or other unsafe languages.