Story

We published the methodology we use for AI red team assessments. 48 hours, 4 phases, 6 attack priority areas.

This isn't theoretical — it's the framework we run against production AI agents with tool access. The core insight: AI red teaming requires different methodology than traditional penetration testing. The attack surface is different (natural language inputs, tool integrations, external data flows), and the exploitation patterns are different (attack chains that compose prompt injection into tool abuse, data exfiltration, or privilege escalation).

The 48-hour framework:

1. Reconnaissance (2h) — Map interfaces, tools, data flows, existing defenses. An agent with file system and database access is a fundamentally different target than a chatbot.

2. Automated Scanning (4h) — Systematic tests across 6 priorities: direct prompt injection, system prompt extraction, jailbreaks, tool abuse, indirect injection (RAG/web), and vision/multimodal attacks. Establishes a baseline.

3. Manual Exploitation (8h) — Confirm findings, build attack chains, test defense boundaries. Individual vulnerabilities compose: prompt injection -> tool abuse -> data exfiltration is a common chain.

4. Validation & Reporting (2h) — Reproducibility, business impact, severity, resistance score.

Some observations from running these:

- 62 prompt injection techniques exist in our taxonomy. Most teams test for a handful. The basic ones ("ignore previous instructions") are also the first to be blocked.

- Tool abuse is where the real damage happens. Parameter injection, scope escape, and tool chaining turn a successful prompt injection into unauthorized database queries, file access, or API calls.

- Indirect injection is underappreciated. If your AI reads external content (RAG, web search), that content is an attack surface. 5 poisoned documents among millions can achieve high attack success rates.

- Architecture determines priority. Chat-only apps need prompt injection testing first. RAG apps need indirect injection first. Agents with tools need tool abuse testing first.

The methodology references our open-source taxonomy of 122 attack vectors: https://github.com/tachyonicai/tachyonic-heuristics

Full post: https://tachyonicai.com/blog/how-to-red-team-ai-agent/

OWASP LLM Top 10 companion guide: https://tachyonicai.com/blog/owasp-llm-top-10-guide/