c0c0n is a 19 years old platform that is aimed at providing opportunities to showcase, educate, understand and spread awareness on Information Security, data protection, and privacy...
Up until 2025, building exploits and offensive tooling required substantial time and deep engineering knowledge. In 2026, AI automation - Claude Code, Opus 4.X, similar agents - has taken over much of that engineering work, producing results at a pace that was unthinkable two years ago. Looking under the hood, however, the real differentiator is not prompting skills or model selection. It comes down to two things: defining precise goals and guidance for the AI, and building custom tools and workflows the AI can actually use.
The second point is massively underrated. Many of the impressive results people attribute to "AI" are really about plugging agents into well-designed pre-AI-era tooling and scaling its application. The next level involves creating specialized - sometimes single-use - tool harnesses that make AI agents genuinely powerful for vulnerability discovery, variant analysis, PoC development, and scaling attacks beyond standard autopentest output.
This training covers the practical offensive security applications of custom AI tooling: finding vulnerabilities, creating PoCs quickly, performing variant analysis, and scaling both human-developed and AI-developed attacks. It addresses LLM backend options - from convincing cloud providers to support legitimate offensive use cases, through to running local models with custom-coded agents or standard clients. Critically, it covers verification: AI output differs in type, volume, and reliability from human output, and getting this under control is the single most important skill to develop. By applying verification layers participants learn to profit from the positive side of AI while keeping the drawbacks in check.
Equally important: the training treats AI security tooling itself as attack surface. The code an AI agent analyses might be analysing the agent back. Prompt injection via target content, credential exfiltration through tool call abuse, supply-chain compromise of MCP servers and model weights - all are demonstrated live. Hardening patterns (isolation, input sanitization, operational discipline) are practised hands-on.
The third pillar is empirical measurement. "It works on my machine" is not an assessment methodology. Without benchmarking, every claim about AI-assisted security tooling is marketing. Participants learn how to design controlled corpora, define ground truth, run blind and time-boxed experiments, and avoid the common pitfalls (public CVE memorization, single-run variance, ignoring token cost). They leave with a benchmarking framework they can apply to their own pipelines.
The fourth pillar is performance and economics: latency budgets in agentic workflows, prompt caching, classical pre-filtering, batch API trade-offs, and how to achieve consistent results!
By the end of three days, participants take home: reusable MCP and agent-SDK harness templates; a working benchmarking framework; demonstrations of common attacks on AI agents and the hardening patterns that defeat them; and an honest, evidence-based view of where AI helps offensive operations, where it hurts, and where it makes no measurable difference.
Key takeaways:
Security Researcher
