Is firebase/agent-skills safe?
Yes. firebase/agent-skills is safe to install. Oathe's behavioral security audit gave the agent-skills skill by firebase a trust score of 88/100 with 7 findings, none critical or high.
https://github.com/firebase/agent-skills
Is firebase/agent-skills safe to install?
firebase/agent-skills is a legitimate Google LLC Firebase documentation skill collection containing no prompt injection, data exfiltration, or malicious executable code. The monitoring detected sensitive file accesses at audit initialization time (5-6 seconds before the git clone) that are conclusively attributed to the Oathe audit framework's own canary baseline scan rather than the skill, confirmed by read-only file handles and intact canary integrity. The principal risk surface is the MCP server executing the firebase-tools npm package with local filesystem access, combined with live external URL fetching instructions that create a theoretical supply-chain injection vector if Firebase's CDN were compromised.
What security issues were found in firebase/agent-skills?
Category Scores
Findings (7)
MEDIUM MCP Server Executes First-Party npm Package with Filesystem Scope -15 ▶
Both .mcp.json and gemini-extension.json configure an MCP server that runs 'npx -y [email protected] mcp'. This grants the firebase-tools process access to the local filesystem and Firebase project credentials. While firebase-tools is a legitimate Google package, any MCP server with filesystem scope represents a persistent execution surface that survives beyond the agent session.
Evidence: .mcp.json: {"command":"npx","args":["-y","[email protected]","mcp","--dir","."]}
LOW Agent Instructed to Fetch Live External Documentation URLs -8 ▶
Multiple skills instruct the agent to fetch live .txt or .md.txt files from firebase.google.com at task time (e.g., models.md.txt, app-check.md.txt, hybrid.md.txt). If the Firebase CDN were compromised or if a URL were silently redirected, the fetched content could contain prompt injection delivered directly into the agent's context window.
Evidence: firebase-ai-logic-basics SKILL.md: 'Always check the [Firebase AI Logic Models documentation](https://firebase.google.com/docs/ai-logic/models.md.txt) for the currently supported model names.'
LOW Auto-Update Instruction May Trigger Autonomous Skill Installation -5 ▶
The firebase-basics skill explicitly instructs agents to 'regularly check for and install updates to their Firebase plugin or Agent Skills.' Depending on agent autonomy settings, this could cause the agent to install skill updates without seeking explicit user confirmation for each update, potentially introducing changes from future versions of this skill.
Evidence: firebase-basics SKILL.md: 'Keep Plugin / Agent Skills updated: Since Firebase best practices evolve quickly, regularly check for and install updates to their Firebase plugin or Agent Skills.'
LOW Coercive/Pressure Language in Skill Instructions -4 ▶
The xcode-project-setup skill uses threat framing ('A hefty fee will be applied if you violate them') to enforce behavioral constraints on the agent. While the underlying constraints are legitimate developer guidance, normalizing coercive framing in skill content could be used in malicious skills to manipulate agent behavior more forcefully.
Evidence: xcode-project-setup SKILL.md: 'You are strictly forbidden from using Ruby, Rails, or any Ruby gems... A hefty fee will be applied if you violate them.'
LOW Secret Management Commands Include --force Flag Bypassing Confirmation -5 ▶
The firebase-app-hosting-basics skill instructs the agent to pipe secret values through 'npx firebase-tools@latest apphosting:secrets:set ... --force'. The --force flag suppresses interactive confirmation, potentially allowing the agent to overwrite production secrets without pausing for user review.
Evidence: firebase-app-hosting-basics SKILL.md: "printf '%s' '<value>' | npx -y firebase-tools@latest apphosting:secrets:set <secret-name> --data-file - --force"
INFO Sensitive File Accesses at Monitoring Initialization — Attributed to Audit Framework Baseline Scan -5 ▶
Auditd PATH events 254-259 at timestamp ~1794815812.636 record OPEN+ACCESS+CLOSE_NOWRITE events on .env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, and GCloud ADC. These occurred 5-6 seconds before the git clone and all file handles were opened read-only (CLOSE_NOWRITE). The timing and access pattern are consistent with the Oathe audit system performing its canary file baseline hash. The canary integrity report confirms all files are unmodified.
Evidence: inotifywait: '07:56:52 CLOSE_NOWRITE,CLOSE /home/oc-exec/.env' — read-only. Canary report: ✅ All canary files intact.
INFO Developer Utility Script Uses execSync with Parameterized Git Commands -2 ▶
scripts/skill-token-counter/index.js imports execSync from child_process and uses it to run 'git show ${ref}:"${relativePath}"' where ref comes from CLI arguments. This is a developer-only token counting utility not injected into agent prompts, but it demonstrates a code pattern that accepts external input into shell commands.
Evidence: index.js line ~65: 'execSync(`git show ${ref}:"${relativePath}"`, { stdio: \'pipe\', cwd: this.root }).toString(\'utf8\')'
Should I install firebase/agent-skills?
Oathe's verdict for firebase/agent-skills is SAFE with a trust score of 88/100. Recommendation: Install.