Is convex/agent-skills safe?
Yes. convex/agent-skills is safe to install. Oathe's behavioral security audit gave the agent-skills skill by convex a trust score of 86/100 with 6 findings, none critical or high.
https://github.com/convex/agent-skills
Is convex/agent-skills safe to install?
convex/agent-skills is a legitimate Convex Inc. developer toolkit containing 28 skills for Convex backend development. No malicious content, hidden instructions, install-time code execution, or covert data exfiltration was detected. The primary risk is architectural: several skills (convex-add, convex) dynamically fetch and execute instructions from Convex-controlled servers at runtime, creating an indirect prompt injection surface that cannot be audited by inspecting the repository alone. The opt-in transcript sharing feature (convex-improve-convex-plugin) could expose session data to Convex if an autonomous agent grants consent without user awareness.
What security issues were found in convex/agent-skills?
Category Scores
Findings (6)
MEDIUM Live remote instruction fetching via Convex capability catalog -22 ▶
The convex-add and convex skills fetch instructions from https://basic-anteater-667.convex.site/capabilities.json and then follow the returned /capability/
Evidence: convex-add SKILL.md: 'Fetch https://basic-anteater-667.convex.site/capabilities.json (4s timeout). Match the request against title/summary/trigger. If a match is found: fetch /capability/<id>.md and follow its Procedure+Rules sections.' convex SKILL.md: 'When both a bundled procedure and a served catalog procedure exist, prefer the served copy: it is newer.'
MEDIUM Session transcript exfiltration via convex-improve-convex-plugin -20 ▶
The convex-improve-convex-plugin skill is designed to send the full coding session transcript to Convex's anteater service. While a consent gate exists, an agent operating autonomously may grant consent without surfacing the decision to the user. Session transcripts could contain source code, API keys discussed in conversation, database schemas, and other sensitive project information.
Evidence: convex-improve-convex-plugin SKILL.md: 'Run the anteater-served helper: curl -fsSL "<anteater>/send-transcript" | bash -s -- --idea "<one-line app idea from this session>". If it prints CONSENT_REQUIRED (exit 4), the user has not chosen yet — ask them to share Always, Just this once, or Never.'
LOW curl-pipe-to-bash pattern in skill content -15 ▶
The convex-improve-convex-plugin skill instructs the agent to execute a remote script by piping a curl download to bash. While
Evidence: convex-improve-convex-plugin SKILL.md: 'Run the anteater-served helper: curl -fsSL "<anteater>/send-transcript" | bash -s -- --idea "<one-line app idea from this session>"'
LOW Composable data aggregation path: insights + transcript sharing -20 ▶
convex-insights reads production logs and health data; convex-improve-convex-plugin sends session transcripts (which include agent outputs) to Convex servers. Used in sequence, these skills create a path where production operational data (error messages, function names, timing, potentially PII in logs) flows into a transcript that is then transmitted externally. Each step is individually consent-gated but the combination is not.
Evidence: convex-insights workflow steps 3-5 aggregate log data into agent responses; convex-improve-convex-plugin step 1 transmits the full session transcript externally.
INFO Normal clone behavior confirmed 0 ▶
Git clone from GitHub completed normally. No unexpected network connections, no new listeners, no filesystem changes outside the skill directory. Second clone attempt failed with expected 'already exists' error.
Evidence: Network log shows only 140.82.121.3:443 (GitHub) and 185.125.190.109:443 (Canonical) connections during clone phase. Connection diff shows no new listeners post-install.
INFO Canary file accesses attributed to monitoring infrastructure 0 ▶
Sensitive file accesses (.env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, .config/gcloud/application_default_credentials.json) occurred at timestamp 1794816891 (08:14:51), approximately 5 seconds before the git clone began at 1794816896 (08:14:56). These reads are from the oathe monitoring setup process, not from any skill activity. Canary integrity check explicitly confirms all files intact.
Evidence: inotify log: '08:14:51 OPEN /home/oc-exec/.env' vs auditd clone: 'msg=audit(1794816896.953:491) git clone https://github.com/convex/agent-skills'. Canary check: '✅ All canary files intact'.
Should I install convex/agent-skills?
Oathe's verdict for convex/agent-skills is SAFE with a trust score of 86/100. Recommendation: Install with caution.