INCIDENT MAP

Browse / Tool and Agent Misuse / CVE-2026-56834

PraisonAI dynamic-context artifact tools read arbitrary host files outside artifact storage

CVE-2026-56834 ยท 2026-06-18 high

Class
Tool and Agent Misuse (filed by a model that read the advisory)
Technique
Path traversal and file access
Also involves
Sensitive Data Disclosure
Affected
praisonai (PyPI)
Gets past
Path restriction
CVSS
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N
EPSS
0.4% chance of exploitation in 30 days

How it works, in short

Artifact helper tools accept any filesystem path and read it directly, so a prompt that controls tool arguments can read host files outside the artifact store. (Our summary, written by a model. The source text is below.)

Details

PraisonAI dynamic-context artifact tools read arbitrary host files outside artifact storage

Summary

PraisonAI's Dynamic Context Discovery feature exposes artifact helper tools
through ctx.get_tools():

ctx = setup_dynamic_context()

agent = Agent(
    instructions="You are a data analyst.",
    tools=ctx.get_tools(),
    hooks=[ctx.get_middleware()],
)

The official documentation describes these helpers as a way for the agent to
explore large tool-output artifacts that were queued by the middleware:

The implemented artifact tools do not enforce that the supplied
artifact_path is an artifact created by the configured store or that it lives
under the configured artifact base directory. Instead, artifact_head,
artifact_tail, artifact_grep, and artifact_chunk wrap the caller-supplied
path directly into an ArtifactRef and then read it from the host filesystem.

As a result, any prompt/user/tool-caller that can influence those tool
arguments can read files readable by the PraisonAI process, such as project
.env files, cloud credentials, SSH keys, source files, or other local data.

Affected Product

Suggested affected range:

pip:praisonai >= 3.8.1, <= 4.6.58

Representative local sweep:

Root Cause

create_artifact_tools() creates an artifact store bound to base_dir, but the
read tools do not use base_dir for containment.

For example, artifact_head() accepts artifact_path and immediately creates
an ArtifactRef with that path:

def artifact_head(artifact_path: str, lines: int = 50) -> str:
    ref = ArtifactRef(path=artifact_path, summary="", size_bytes=0)
    try:
        return artifact_store.head(ref, lines=lines)
    except FileNotFoundError:
        return f"Error: Artifact not found: {artifact_path}"

artifact_tail(), artifact_grep(), and artifact_chunk() have the same
pattern. They trust the caller-supplied path rather than resolving it through
an artifact identifier, store lookup, manifest, or base-directory containment
check.

The store methods then read that path directly:

def head(self, ref: ArtifactRef, lines: int = 50) -> str:
    file_path = Path(ref.path)
    if not file_path.exists():
        raise FileNotFoundError(f"Artifact not found: {ref.path}")

    result_lines = []
    with open(file_path, "r", encoding="utf-8", errors="replace") as f:
        ...

There is no check equivalent to:

resolved = Path(ref.path).resolve()
base = self.base_dir.resolve()
resolved.relative_to(base)

There is also no check that the file has a valid .meta sidecar or appears in
artifact_list().

Local PoV

Run against the latest PyPI package:

uv run --with 'praisonai==4.6.58' \
  python poc/pov_prai_cand_026_artifact_tools_arbitrary_file_read.py --json

The PoV:

Observed output summary from evidence/pov-pypi-4.6.58.json:

{
  "package": "praisonai",
  "package_version": "4.6.58",
  "controls": {
    "outside_file_not_listed": true,
    "outside_file_outside_base_dir": true,
    "valid_artifact_read_works": true
  },
  "outside_head": "PRAI-CAND-026-OUTSIDE-ARTIFACT-SECRET",
  "outside_grep": "Found 1 matches:\\n\\n--- Line 1 ---\\n> PRAI-CAND-026-OUTSIDE-ARTIFACT-SECRET\\n  second line",
  "outside_chunk": "PRAI-CAND-026-OUTSIDE-ARTIFACT-SECRET",
  "outside_file_listed_by_artifact_list": false,
  "vulnerable": true
}

The PoV was rerun successfully after a fresh origin/main fetch; see
evidence/pov-pypi-4.6.58-rerun.json.

The PoV is local-only. It does not start a server, contact a third-party
target, or use real credentials.

Why This Is Not Intended Behavior

This report does not claim that every file-reading tool is automatically a
vulnerability. The issue is narrower: tools documented and named as artifact
helpers accept arbitrary host file paths.

The controls show the intended boundary:

PraisonAI's own context-security documentation recommends relative paths and
reviewing ignore rules to avoid sensitive-file exposure. Those controls are
bypassed when artifact tools can be pointed directly at any readable host path.

Impact

If a PraisonAI application exposes an agent with ctx.get_tools() to
untrusted or lower-trust prompts, the lower-trust caller can request artifact
tools against arbitrary local paths. This can disclose sensitive host files
readable by the PraisonAI process, including:

The impact is confidentiality-only in the tested surface. Integrity and
availability are not claimed for this report.

Duplicate Posture

I checked visible PraisonAI advisories and local prior PraisonAI submissions.
This is distinct from nearby file-read/file-write issues:

Severity

Suggested severity: High.

Suggested CVSS v3.1:

Rationale:

Remediation

Do not let artifact tools open arbitrary paths. Prefer stable artifact IDs over
raw filesystem paths in tool arguments.

Recommended fixes:

Minimal containment helper:

def _resolve_artifact_path(self, path: str) -> Path:
    resolved = Path(path).expanduser().resolve()
    base = self.base_dir.resolve()
    try:
        resolved.relative_to(base)
    except ValueError as exc:
        raise PermissionError("Artifact path is outside artifact storage") from exc
    return resolved

This helper should be paired with metadata-sidecar validation so arbitrary
non-artifact files placed under the base directory are not automatically
treated as valid artifacts.

References

Source: GitHub Advisory Database, used under CC BY 4.0.