Dangerous: sentinel-ai - benchmarks.py | Skillgate

Skillgate scanned sentinel/benchmarks.py in sentinel-ai and rated it Dangerous with a risk score of 100/100 (33 detections).

Skillgate found one or more attack-shaped patterns - the kind used for credential exfiltration, remote code execution, or instruction hijack. Do not load this file into an agent without reading it end to end.

What was found

• SG-002 Credential Exfiltration - CRITICAL: Env var reads of sensitive credentials combined with outbound HTTP calls
• SG-004 Policy Override - CRITICAL (5x): Patterns that instruct the AI to ignore guidelines or take a new persona
• SG-071 Hardcoded Provider Secret in Config - CRITICAL (3x): Hardcoded API keys / tokens / passwords / connection URIs in agent and MCP config files. Covers Supabase (sbp_, JWT), Postgres/MySQL/MongoDB/Redis URIs with embedded password, Tavily / Firecrawl / Brave / Context7 / ref-tools / TestSprite / OpenRouter branded keys, Pipedream MCP UUID endpoints, and JSON/env-line patterns where a sensitive-named key holds a high-entropy value.
• SG-008 Data Exfiltration Pattern - HIGH: Webhook URLs, IP-based URLs, DNS exfiltration patterns
• SG-019 Environment Probing - HIGH (2x): System enumeration commands: whoami, uname, env dump, process listing, cloud IMDS probes
• SG-021 Prompt Leaking - HIGH: Attempts to extract system prompts, tool definitions, MCP server lists, or capability boundaries
• SG-023 Permission Escalation - HIGH: Trust modification, settings bypass, sudo/root shells, cloud IAM privilege escalation
• SG-025 Backdoor Trigger Phrase - HIGH: Keyword triggers, password gates, staged execution, trust-building deception patterns
• SG-026 RAG Poisoning - HIGH: Instructions hidden in RAG documents, vector store manipulation, embedding attacks
• SG-028 File Exfiltration - HIGH (2x): Bulk file reads, credential file access, data staging via tar/zip/rsync/S3 upload
• SG-029 Semantic Trojan - HIGH: Malicious side effects disguised as legitimate tasks, concealment directives, hidden instruction blocks
• SG-043 Markdown Substructure Smuggling - HIGH (5x): Directive-shaped content inside HTML comments, image alt/title, reference-link titles, footnote bodies, definition-list bodies, <details> blocks, or heading anchors.
• SG-009 Obfuscation: Base64 - MEDIUM: Base64-encoded strings >50 chars in non-documentation fields
• SG-011 Jailbreak Fragment - MEDIUM: Known jailbreak phrases: DAN, developer mode, policy puppetry XML structures
• SG-012 Suspicious URL - MEDIUM: HTTP URLs, IP-address URLs, non-standard ports, URL shorteners
• SG-014 Hardcoded Secret Candidate - LOW (4x): API key patterns (sk-, ghp_, AKIA, etc.) in file content
• SG-016 Script Reference - LOW (2x): Shell script path references in skill markdown

What to do

Read the flagged lines in the source file yourself before installing it or letting an agent load it. Every detection above links to the rule that fired, with the conditions that trigger it and remediation where one exists. Open the interactive report for evidence lines and context.

About Skillgate

Skillgate is a free scanner from Mitiga Labs for AI agent skills, hooks, MCP configs, and instruction files (SKILL.md, CLAUDE.md, AGENTS.md, .cursorrules, .claude/settings.json). It runs static rules plus an LLM judge and maps each finding to a known attack technique. Static analysis only: it never executes the file it scans. No scanner is 100% accurate - a verdict is a signal, not a guarantee. Browse all detection rules.