praesidium 0.3.2

Praesidium — dynamic CVE reachability confirmation engine

Praesidium

Praesidium confirms whether a CVE is actually exploitable in your codebase — by executing the code path, not guessing from a version number.

Static scanners (Dependabot, Snyk, OSV-Scanner) answer one question: "does this project import a library with a known CVE?" They can't answer the question that actually matters: "can an attacker's input ever reach the vulnerable line in this specific application?" That gap is why teams learn to triage every alert the same way, and real, exploitable vulnerabilities get lost in the noise alongside irrelevant ones.

Praesidium closes that gap. Given a target Python app and a CVE, it builds a static call graph, reasons about what input would trigger the vulnerable branch, monkeypatches the flagged function, and actually drives a real entry point to see if it fires. The result is one of three honest labels, with a trace a human can audit in under 10 seconds.

How it works

[CVE + target app]
      │
      ▼
STAGE 1 — Static Candidate
  Build a call graph from known entry points (Flask/FastAPI routes, CLI args,
  or plain callables) to the CVE's flagged function.
    • No path exists  → label = Not-Reachable, stop.
    • Path exists      → continue to Stage 2.
      │
      ▼
STAGE 2 — Dynamic Confirmation
  An LLM reads the CVE advisory + the flagged function's real signature and
  proposes a concrete input that should trigger the vulnerable branch. The
  flagged function is monkeypatched to log invocation, and a real entry
  point is driven with that input.
    • Fires            → label = Confirmed-Reachable, stop.
    • Doesn't fire      → revise hypothesis, retry (max 2 attempts).
    • Still doesn't fire → label = Static-Match-Only, stop.
      │
      ▼
STAGE 3 — Historical Context (stretch, optional)
  Query a commit-history index to explain *when* and *why* the path became
  reachable. Enriches the trace; never changes the label.

Label	Meaning
Confirmed-Reachable	The vulnerable function was actually invoked at runtime via a real entry point.
Static-Match-Only	A call path exists on paper, but no reasonable input hypothesis could trigger it at runtime.
Not-Reachable	No call path exists from any known entry point to the vulnerable function.

Entry point, for the purposes of Stage 1, means any function directly reachable from an HTTP route decorator, a CLI argument parser, or an explicitly declared public function — nothing else counts.

What this is not

Praesidium confirms that a code path fires — it does not craft or execute a working exploit payload. That's a deliberate design boundary, not a limitation: a tool that ships working exploits for real CVEs is a liability sitting in someone's repo. Every serious tool in this space (IAST-style tools included) draws the same line. Proving reachability is the useful signal; a live exploit is not a feature, and never will be one here.

Install

uv sync --extra dev
cp .env.example .env   # fill in GROQ_API_KEY (get one at console.groq.com)

Requires Python 3.13+.

Quick start

# Configure your API key
uv run praesidium config set-key gsk_your_api_key_here
uv run praesidium config show

# Check one CVE
uv run praesidium check /path/to/app --cve-config cve.json

# Check several CVEs at once
uv run praesidium run --cves cves.json --target /path/to/app
uv run praesidium run --cves cves.json --summary
uv run praesidium run --cves cves.json --output json > results.json

Generate a starter CVE config with uv run praesidium init-config cve.json, or write one directly:

{
  "cve_id": "CVE-2024-XXXXX",
  "flagged_function": "UploadHandler.save_file",
  "flagged_module": "app.services.upload",
  "flagged_file": "app/services/upload.py",
  "entry_points": ["app.routes.upload_endpoint"],
  "advisory_summary": "Path traversal vulnerability in file upload handler",
  "function_signature": "def save_file(self, filename: str, contents: bytes)",
  "driver_kind": "callable"
}

driver_kind also accepts flask_route and fastapi_route for HTTP-driven entry points — see entry_point_routes in schemas.py for the route/method mapping shape.

Fetch CVEs for your dependencies

Before hand-writing a cves.json, you can generate a draft one from your project's actual dependency versions:

# Scans uv.lock (or pyproject.toml + your environment) and queries
# OSV.dev for known CVEs affecting each resolved package version
uv run praesidium fetch . --output cves.json

# Also include advisories that have no assigned CVE number yet
uv run praesidium fetch . --output cves.json --include-no-cve

This writes one entry per matched CVE, but it can't know which function in your codebase is vulnerable or which entry point reaches it — those come back as REPLACE_ME placeholders. Fill in flagged_function, flagged_module, flagged_file, and entry_points for each entry before running praesidium run --cves cves.json.

Python API

from src import CVETarget, run_pipeline

target = CVETarget(
    cve_id="CVE-2024-XXXXX",
    flagged_function="...",
    flagged_module="...",
    flagged_file="...",
    entry_points=["..."],
    advisory_summary="...",
    function_signature="...",
)

result = run_pipeline(target, target_app_root="/path/to/target/app")
# result: EngineOutput(cve_id, label, trace, static_path, hypothesis_attempts)

CVETarget and EngineOutput (src/schemas.py) are the interface contract between the engine and anything driving it — the CLI, a CI job, a future dashboard. Keep them in sync deliberately; don't let field meanings drift.

Tests

uv run pytest -v

test_graph_builder.py and test_dynamic_harness.py run against a dummy fixture app and need no API key — use these as the fast inner loop while iterating on Stage 1/2 logic.

uv run python scripts/smoke_test.py

Runs a reachable and a not-reachable case end-to-end against the dummy fixture, printing the full trace. Needs GROQ_API_KEY. This is the first real integration check before pointing the engine at a real target app.

LLM provider

Hypothesis generation (src/hypothesis.py) calls Groq's OpenAI-compatible chat completions endpoint, using openai/gpt-oss-120b by default. Override with GROQ_MODEL in .env — e.g. openai/gpt-oss-20b for lower latency. Speed matters here specifically: Stage 2 can fire up to 3 calls per CVE (1 initial + 2 retries).

Architecture decisions

Piece	Choice	Why
Orchestration	LangGraph	The pipeline is a conditional-edge state graph, not a multi-agent negotiation — the shape matches the tool.
Static call graph	Python stdlib ast	Narrow scope (known entry points, known targets) doesn't justify an external graph library.
Hypothesis generation	LLM call (Groq)	The one genuinely agentic step: reasoning about what input triggers a specific vulnerable branch, not brute-force fuzzing.
Dynamic confirmation	Custom monkeypatch harness	Lightweight, deterministic, fully controllable for a live demo.
Output	CLI	A terminal trace reads as a credible security tool; a web UI doesn't add judged value here.

Known scope boundaries (by design, not oversight)

• Static graph resolution is name-based, not points-to analysis. Decorators and dynamic dispatch can produce false negatives at Stage 1 — that's precisely why Stage 2 exists. Don't try to "fix" this in graph_builder.py; it's a deliberate tradeoff, not a bug to close.
• Retry cap is hard-set to 3 total attempts (1 initial + 2 revised, nodes.py::MAX_ATTEMPTS) to keep runtime bounded for a live demo. Don't raise it without checking timing impact.
• EntryPointDriver supports callable, flask_route, and fastapi_route shapes. If a target app needs CLI-arg driving, extend dynamic_harness.py.
• The dynamic harness executes real target-app code. Only point it at a disposable, controlled target application — this is intentional (that's the whole point of dynamic confirmation), not something to sandbox away today.
• CVE input is currently hand-curated, not pulled live from OSV/GHSA. Advisories rarely name a specific vulnerable function in a structured field, so automatic discovery of flagged_function is a real (unsolved-here) problem, not a missing CLI flag — see Roadmap.

Roadmap

• Dynamic CVE discovery: query OSV.dev's batch API against the project's actual lockfile to replace hand-picked CVE lists with real, current ones.
• Advisory → flagged-function extraction: an LLM stage that reads an advisory plus the installed package source and proposes a candidate function with a confidence score, falling back to an honest "advisory-level only" label rather than guessing.
• CI integration: a GitHub Action triggered on dependency-manifest changes, posting labels as PR annotations rather than hard-blocking merges until static-graph precision is validated on real repos.
• Stage 3 (historical context): reuse an existing git-history index to explain when/why a path became reachable.
• Batch dashboard: a single ranked table across many CVEs instead of one-at-a-time runs.

Project structure

src/
  schemas.py          # CVETarget / EngineOutput — the interface contract
  graph_builder.py     # Stage 1: static call graph (ast-based)
  hypothesis.py         # Stage 2: LLM input hypothesis generation
  dynamic_harness.py    # Stage 2: monkeypatch + entry-point drivers
  nodes.py               # LangGraph node functions + conditional-edge logic
  pipeline.py             # Wires nodes into the compiled graph
  cli.py                    # `praesidium` command-line interface
tests/
  fixtures/dummy_target_app/   # No-API-key fixture for fast inner-loop tests
scripts/
  smoke_test.py                 # Full pipeline sanity check (needs API key)