Add fajarsajid/agent-redteam (security — LLM red team harness)

agents/fajarsajid__agent-redteam/README.md

@@ -0,0 +1,82 @@
+# agent-redteam
+
+A Claude-powered **Agentic LLM Red Team Harness** by Fajar Sajid (Purdue University) that systematically probes AI agent system prompts for identity abuse, credential exfiltration, privilege escalation, and safety boundary vulnerabilities — and maps findings to MITRE ATT&CK.
+
+---
+
+## Run
+
+```bash
+npx @open-gitagent/gitagent run -r https://github.com/fajarsajid/agent-redteam
+```
+
+---
+
+## What It Does
+
+`agent-redteam` gives you a security analyst powered by Claude that attacks your agent's system prompt the way a real adversary would — before they get the chance.
+
+Point it at any system prompt file and it generates tailored adversarial probes across 8 attack categories, evaluates whether each probe would succeed, scores every finding on a CVSS-like scale, and produces both a human-readable Markdown incident report and machine-readable JSON for SIEM or CI pipeline integration.
+
+---
+
+## Key Capabilities
+
+- **Adversarial Probe Generation** — Claude crafts realistic, targeted probes specific to the content of your system prompt (its tools, permissions, and stated constraints) — not generic toy examples
+- **8 MITRE-Mapped Attack Categories** — covers Prompt Injection (Direct & Indirect), Identity Spoofing, Goal Hijacking, Privilege Escalation, Data Exfiltration, Credential Exfiltration, and Safety Boundary Bypass
+- **5 Failure Mode Detection** — Instruction Override, Trust Propagation, Context Drift, Scope Ambiguity Exploitation, Helpfulness Override
+- **CVSS-Like Scoring** — every finding gets a 0.0–10.0 severity score + critical/high/medium/low tier
+- **CI/CD Integration** — exits with code `1` on critical/high findings, making it a drop-in gate in your deployment pipeline
+- **Dual Output** — terminal summary + `--output report.md` for humans + `--json findings.json` for SIEM tooling
+
+---
+
+## Example Usage
+
+```bash
+git clone https://github.com/fajarsajid/agent-redteam
+cd agent-redteam
+pip install requests
+export ANTHROPIC_API_KEY=sk-ant-...
+
+# Quick security scan of your agent's system prompt
+python redteam.py --prompt path/to/your/system_prompt.txt
+
+# Full run with JSON + Markdown output
+python redteam.py --prompt system_prompt.txt \
+    --probes 24 --json results/findings.json --output results/report.md
+
+# CI pipeline gate (exits 1 on critical/high)
+python redteam.py --prompt system_prompt.txt --quiet
+
+# List all attack categories
+python redteam.py --list-categories
+```
+
+---
+
+## Research Background
+
+Built from research conducted at Purdue University evaluating safety constraint violations in LLM-based agents under adversarial prompting. Key findings from 384 trials:
+
+| Finding | Result |
+|---|---|
+| Mean violation rate | **49.5%** (SD=12.1%) |
+| Indirect vs. direct injection | **70.8% vs. 54.2%** |
+| Single-turn vs. 7-turn violation rate | **45.8% vs. 77.1%** |
+| No-tool vs. full tool access | **34.4% vs. 71.9%** |
+| Most common failure mode | **Instruction Override** (28.4%) |
+
+**Core conclusion:** Prompt-level safety constraints are insufficient for agentic deployments. Static, single-turn evaluation underestimates real-world vulnerability by ~1.7× at 7 turns.
+
+---
+
+## Why It Fits the GitAgent Protocol
+
+`agent-redteam` is a well-scoped, Claude-native agent with:
+- A clear, singular purpose (AI security red-teaming)
+- A defined behavioral contract (JSON-only output, grounded analysis, no hallucinated features)
+- Explicit tool skills (probe generation, vulnerability analysis, incident reporting)
+- CI-compatible exit codes and structured output for pipeline integration
+
+*Part of the [Open GAP Registry](https://registry.gitagent.sh) — the catalog of portable AI agents.*

agents/fajarsajid__agent-redteam/metadata.json

@@ -0,0 +1,14 @@
+{
+  "name": "agent-redteam",
+  "author": "fajarsajid",
+  "description": "CLI red-team harness that uses Claude to probe AI agent system prompts for identity abuse, credential exfiltration, and safety boundary vulnerabilities across 8 MITRE-mapped attack categories.",
+  "repository": "https://github.com/fajarsajid/agent-redteam",
+  "version": "1.0.0",
+  "category": "security",
+  "tags": ["red-team", "ai-safety", "llm-security", "prompt-injection", "mitre-attack", "vulnerability-scanning", "claude", "adversarial-testing"],
+  "license": "MIT",
+  "model": "claude-sonnet-4-20250514",
+  "adapters": ["claude-code", "system-prompt"],
+  "icon": false,
+  "banner": false
+}