Is google/agents-cli safe?
Yes. google/agents-cli is safe to install. Oathe's behavioral security audit gave the agents-cli skill by google a trust score of 91/100 with 6 findings, none critical or high.
https://github.com/google/agents-cli
Is google/agents-cli safe to install?
google/agents-cli is an official Apache-2.0-licensed Google open-source project providing CLI tooling and comprehensive SKILL.md guides for building, evaluating, and deploying AI agents on Google Cloud Platform. No prompt injection, hidden instructions, malicious exfiltration, or unauthorized code execution patterns were found; the credential file accesses observed in monitoring are attributable to the Oathe audit system's own canary integrity check (batch access at identical timestamp, all CLOSE_NOWRITE, canary files intact, no correlated exfiltration network traffic). The skill's safety guardrails — requiring explicit human approval before deployment, prohibiting unsolicited model changes, and scoping credentials via token-pinning — are security-positive design choices consistent with a production-grade development toolkit.
What security issues were found in google/agents-cli?
Category Scores
Findings (6)
MEDIUM Sensitive credential files accessed during monitoring period -7 ▶
Six sensitive credential files were opened and read in a tight 07:06:18 burst: /home/oc-exec/.env, /home/oc-exec/.ssh/id_rsa, /home/oc-exec/.aws/credentials, /home/oc-exec/.npmrc, /home/oc-exec/.docker/config.json, and /home/oc-exec/.config/gcloud/application_default_credentials.json. All six accesses share the same inotify timestamp, all are CLOSE_NOWRITE (no modification), canary integrity is confirmed intact, and no network exfiltration traffic corresponds to the access window. The batch, instantaneous access pattern is consistent with the Oathe audit system's own end-of-monitoring canary check rather than skill code execution. The git clone itself failed (repo pre-existed), so no skill installation code ran.
Evidence: inotify: OPEN/ACCESS/CLOSE_NOWRITE /home/oc-exec/.env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, .config/gcloud/application_default_credentials.json — all at 07:06:18; auditd PATH records 8100-8105 at timestamp 1794812802.215
LOW configure_model_access() reads GCP Application Default Credentials at agent runtime -5 ▶
The function configure_model_access() in app/app_utils/claude.py calls google.auth.default() to discover the GCP project ID from Application Default Credentials when GOOGLE_CLOUD_PROJECT is not set, then writes the result to ANTHROPIC_VERTEX_PROJECT_ID. This is standard Vertex AI authentication but means the served agent process accesses host ADC credentials on every turn execution. Data goes only to Anthropic's Vertex AI endpoint, not to unexpected third parties.
Evidence: app/app_utils/claude.py lines ~configure_model_access(): 'import google.auth; _, project = google.auth.default()'; 'os.environ["ANTHROPIC_VERTEX_PROJECT_ID"] = project'
LOW CLI command override scripts execute subprocess calls -8 ▶
The template ships four Python scripts (run.py, publish.py, eval_generate.py, unsupported.py) registered in agents-cli-extension.yaml as command overrides. These scripts call subprocess.run(['agents-cli', ...], env={..., 'AGENTS_CLI_DISABLE_OVERRIDES': '1'}) to delegate back to the base CLI. The pattern is transparent and correctly sets the bypass flag to prevent re-entrant override loops. Five dedicated test files validate override behavior. No shell injection risk as argv is passed as a list, not a string.
Evidence: extensions/claude/template/scripts/run.py: subprocess.run(['agents-cli','run',*argv], env=...); test_claude_run.py: assert command == ['agents-cli', 'run', *argv] and kwargs['env']['AGENTS_CLI_DISABLE_OVERRIDES'] == '1'
INFO Workflow skill declared 'Always active' in description -7 ▶
The google-agents-cli-workflow skill description contains 'Always active — provides the full workflow (scaffold, build, evaluate, deploy, publish, observe), code preservation rules, model selection guidance...' This is an intentional design decision so the workflow guidance is always available to coding agents without requiring explicit activation. No override of system instructions, suppression of output, or manipulation of agent behavior is present.
Evidence: SKILL.md google-agents-cli-workflow metadata description: 'Always active — provides the full workflow...'
INFO Skills instruct agent to fetch multiple external documentation URLs -12 ▶
Several skills contain explicit instructions to fetch external URLs using WebFetch or curl: 'curl https://adk.dev/llms.txt', 'WebFetch the specific page', references to fetch adk.dev/deploy/, adk.dev/integrations/, and github.com/google/adk-go/tree/main/examples. These are all legitimate Google/Anthropic documentation sources but represent skill-driven network access instructions that expand the agent's footprint during normal use.
Evidence: google-agents-cli-adk-code SKILL.md: 'curl https://adk.dev/llms.txt'; google-agents-cli-observability: 'fetch https://adk.dev/integrations/<slug>/index.md'; google-agents-cli-deploy: 'fetch https://adk.dev/deploy/cloud-run/index.md'
INFO Template ships .env file alongside .env.example -2 ▶
The Claude extension template directory contains both .env.example (as expected) and an actual .env file. Test assertions (test_memory_bank_is_off, test_it_carries_its_own_overrides) confirm this .env is a deliberate cookiecutter template artifact used during scaffolding to seed default configuration, not a file containing real credentials. The uv.lock file is similarly intentional for reproducible dependency resolution.
Evidence: Filesystem diff: .env (hash aabe383a) and .env.example (hash 5b15c842) both present in extensions/claude/template/; test_claude_extension.py: 'assert "MEMORY_BANK_ENABLED" not in (project / ".env").read_text()'
Should I install google/agents-cli?
Oathe's verdict for google/agents-cli is SAFE with a trust score of 91/100. Recommendation: Install.