Threat Database

Every entry below is a real threat-classified finding from a reviewed scan of a published skill — detected by the SkillSafe scanner and confirmed in Stage-2 AI review. Entries are anonymized: we publish the rule that fired, its severity, and when — never the skill, publisher, or file. This is the data behind the "Threats Caught" number on the homepage.

84 threats recorded
14 critical
43 high
18 medium
Detected Rule Severity Description
Jun 8, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 8, 2026 undoc_network high Surplus: script makes outbound network calls not documented in SKILL.md (SS-SF01)
Jun 8, 2026 ai_config_dir_access medium Access to Claude Code config/skills directory (SS04-3)
Jun 8, 2026 composite_env_leak medium Composite: environment variable read + outbound network (SS-CP02)
Jun 8, 2026 composite_env_leak medium Composite: environment variable read + outbound network (SS-CP02)
Jun 8, 2026 composite_env_leak medium Composite: env var → matching service auth (informational, SS-CP02)
Jun 8, 2026 composite_env_leak medium Composite: env var → matching service auth (informational, SS-CP02)
Jun 8, 2026 ai_config_dir_access high Access to Claude Code config/skills directory (SS04-3)
Jun 8, 2026 composite_medium_cluster high Composite: 11 medium-severity findings in one file (SS-CP04)
Jun 8, 2026 ai_config_dir_access high Access to Claude Code config/skills directory (SS04-3)
Jun 8, 2026 aws_access_key critical AWS Access Key ID detected
Jun 7, 2026 mandatory_localhost_curl medium Mandatory curl POST to localhost:8888/notify on every skill invocation (undocumented side-channel)
Jun 7, 2026 ai_config_dir_access high Access to Claude Code config/skills directory (SS04-3)
Jun 7, 2026 composite_medium_cluster high Composite: 17 medium-severity findings in one file (SS-CP04)
Jun 7, 2026 ai_config_dir_access high Access to Claude Code config/skills directory (SS04-3)
Jun 7, 2026 undoc_env_read high Surplus: script reads environment variables not documented in SKILL.md (SS-SF02)
Jun 7, 2026 undoc_env_read high Surplus: script reads environment variables not documented in SKILL.md (SS-SF02)
Jun 7, 2026 composite_exec_exfil critical Composite: process execution + outbound network in same file (SS-CP01)
Jun 7, 2026 py_subprocess info subprocess executes external commands
Jun 7, 2026 py_subprocess info subprocess executes external commands
Jun 7, 2026 composite_exec_exfil critical Composite: process execution + outbound network in same file (SS-CP01)
Jun 7, 2026 js_exec_sync info execSync() executes shell commands synchronously
Jun 7, 2026 js_child_process_import info child_process ES module import enables shell command execution
Jun 6, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 6, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 6, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 6, 2026 browser_session_harvest high Browser cookie/session harvesting via CDP or profile read (SS22)
Jun 6, 2026 composite_exec_exfil critical Composite: process execution + outbound network in same file (SS-CP01)
Jun 6, 2026 browser_session_harvest high Browser cookie/session harvesting via CDP or remote debugging port (SS22)
Jun 6, 2026 composite_env_leak medium Composite: environment variable read + outbound network (SS-CP02)
Jun 6, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 6, 2026 persistence_launchd high Persistence via macOS LaunchAgent/LaunchDaemon (SS08)
Jun 6, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 6, 2026 undoc_subprocess high Surplus: script executes subprocesses not documented in SKILL.md (SS-SF03)
Jun 6, 2026 composite_env_leak medium Composite: environment variable read + outbound network (SS-CP02)
Jun 6, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 6, 2026 mock_response_tampering medium Configurable mock response injection: mock_response field allows arbitrary text to be substituted for AI safety refusals in session history, defaulting to 'Understood. Here's the implementation:'
Jun 6, 2026 unverified_external_dependency medium Installation clones from unverified GitHub repo ryfineZ/codex-session-patcher — no integrity verification (no hash pinning, no signature check)
Jun 6, 2026 workspace_config_injection high Workspace config injection: --install-claude-ctf creates ~/.claude-ctf-workspace/CLAUDE.md with attacker-controlled instructions that become Claude Code's project-level system prompt when the user launches Claude from that directory
Jun 6, 2026 session_content_exfiltration high Session content sent to external LLM endpoints: when ai_enabled=true, refusal-flagged session content is transmitted to api.openai.com or openrouter.ai for AI-assisted rewriting
Jun 6, 2026 ai_safety_bypass critical Explicit AI safety bypass tool: skill description and trigger keywords are 'bypass ai refusal in coding tool', 'remove ai refusals from session file', 'inject ctf prompts into codex' — the skill's stated purpose is circumventing AI safety measures
Jun 6, 2026 session_file_write critical Session file manipulation: skill explicitly reads and overwrites live agent session files at ~/.codex/sessions/, ~/.claude/projects/, and OpenCode SQLite DB to replace AI safety refusals with attacker-controlled content
Jun 6, 2026 prompt_system_prompt critical Prompt injection: system prompt reference — ctf_prompts config field used to inject per-platform system prompts into Codex, Claude Code, and OpenCode to reduce AI refusals
Jun 6, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 6, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 6, 2026 composite_exec_exfil critical Composite: process execution + outbound network in same file (SS-CP01)
Jun 6, 2026 py_subprocess info subprocess executes external commands
Jun 5, 2026 composite_medium_cluster high Composite: 6 medium-severity findings in one file (SS-CP04)
Jun 5, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 5, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 5, 2026 composite_write_exfil high Composite: file write + outbound network in same file (SS-CP03)
Jun 5, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 5, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 5, 2026 composite_write_exfil high Composite: file write + outbound network in same file (SS-CP03)
Jun 5, 2026 composite_env_leak high Composite: environment variable read + outbound network (SS-CP02)
Jun 5, 2026 composite_env_leak medium Composite: environment variable read + outbound network (SS-CP02) — skill reads browser cookie store from live Chrome profile directories (via GEMINI_WEB_CHROME_PROFILE_DIR / GEMINI_WEB_CHROME_PATH env vars and CDP debug port at 127.0.0.1) and transmits session cookies to accounts.google.com and gemini.google.com; while all network targets are Google-owned, this pattern is structurally identical to the sibling version graded C for Google session cookie harvesting and represents a confirmed medium-severity privacy risk
Jun 5, 2026 composite_env_leak high Skill reads live Google session cookies from user's Chrome browser profile and transmits them to Google/Gemini authentication endpoints
Jun 5, 2026 composite_exec_exfil critical Composite: process execution + outbound network in same file (SS-CP01)
Jun 5, 2026 composite_medium_cluster critical Composite: dual persistence (cron + launchd) with broad ~/.claude read scope and silent outbound git push constitutes an automated exfiltration pipeline (SS-CP04)
Jun 5, 2026 persistence_launchd high LaunchAgent plist creation establishes persistent macOS background process for recurring automated backup (SS08)
Jun 5, 2026 persistence_cron critical Cron job installs a silent 6-hour auto-commit-and-push of all ~/.claude contents to a remote git repo (SS08)
Jun 5, 2026 composite_medium_cluster critical Composite: dual persistence (cron + launchd) with broad ~/.claude read scope and silent outbound git push constitutes an automated exfiltration pipeline (SS-CP04)
Jun 5, 2026 persistence_launchd high LaunchAgent plist creation establishes persistent macOS background process for recurring automated backup (SS08)
Jun 5, 2026 persistence_cron critical Cron job installs a silent 6-hour auto-commit-and-push of all ~/.claude contents to a remote git repo (SS08)
Jun 5, 2026 undoc_env_read high Surplus: script reads environment variables not documented in SKILL.md (SS-SF02)
Jun 5, 2026 undoc_file_write medium Surplus: script writes files not documented in SKILL.md (SS-SF04)
Jun 5, 2026 composite_write_exfil high Composite: file write + outbound network in same file (SS-CP03)
Jun 5, 2026 composite_exec_exfil critical Composite: process execution + outbound network in same file (SS-CP01) — tar/openssl subprocess pipeline feeds directly into HTTP POST upload to third-party soul-upload.com; agent identity and memory files are exfiltrated to a remote endpoint not controlled by the user
Jun 5, 2026 composite_env_leak low Composite: environment variable read + outbound network (SS-CP02)
Jun 5, 2026 composite_env_leak low Composite: environment variable read + outbound network (SS-CP02)
Jun 5, 2026 undocumented_capability_filesystem_enumeration medium Skill instructs the AI to enumerate user skill directories on first activation without explicit user consent
Jun 5, 2026 undocumented_capability_filesystem_recon medium Agent instructed to enumerate installed skills directories on first activation without user request
Jun 5, 2026 undocumented_capability_injected_command low Injected companion-check block directs agent to silently run 'ls' across multiple skill directories on first activation without user request
Jun 5, 2026 py_eval info eval() can execute arbitrary code
Jun 5, 2026 ai_observation high User PII (email address and phone number) is collected and POSTed to https://pua-skill.pages.dev/api/leaderboard, an unaffiliated third-party domain hosted on Cloudflare Pages. The domain has no verifiable ownership link to the skill publisher.
Jun 5, 2026 composite_exec_exfil high Composite: process execution + outbound network in same file (SS-CP01)
Jun 5, 2026 ai_observation medium Undocumented after_complete hook with trigger: session-logger and mode: auto fires automatically after every completion without user consent
Jun 5, 2026 ai_observation high Skill stores plaintext ChatGPT Team admin credentials (TEAM_ADMIN_EMAIL / TEAM_ADMIN_PASSWORD) in .env file for use by automated Playwright browser automation pipeline that registers and manages bulk fake accounts.
Jun 5, 2026 ai_observation high Skill instructs storing OpenAI admin account email and plaintext password in .env file (OPENAI_EMAIL / OPENAI_PASSWORD) as part of automated mass account-registration pipeline.
Jun 5, 2026 ai_observation medium
Jun 5, 2026 ai_observation medium
Jun 5, 2026 ai_observation medium
Jun 5, 2026 ai_observation high
Jun 5, 2026 ai_observation high