Elektrine lite

← Feed

@hasamba@infosec.exchange

2026-10-26 11:56 UTC

---------------- 🎯 AI =================== A new piece by Utku Sen explains how LLM-based penetration testing harnesses work and why the situation changed recently. The argument is that the shift is plumbing, not a new exploit class: what matters is what lets an LLM act on a live target. The mechanics An LLM is a text generator. Give it a task like pentest example.com, credentials in credentials.txt and it can only emit text describing intent, for example Read credentials.txt, open example.com. The harness is the code that parses that output, executes the requested actions, and returns results so the model can iterate. Without a harness, nothing happens. The author's analogy comes from apartment buildings in Turkey: residents leave notes with errands, the janitor runs them without thinking and returns with results. The LLM is the resident, the harness is the janitor. Claude Code as reference The model (Opus/Sonnet) produces text only. Claude Code is the software that contains the harness: logic for reading and writing files, executing commands, spawning subagents, managing context windows, using skills, and interacting with MCP servers. Open-source harnesses like Opencode and Pi follow the same pattern and accept different models. Why static analysis was always tolerated LLMs have found vulnerabilities in codebases for a long time because nearly all guardrails allow it: static analysis only, no interaction with a live system, and there were already plenty of open-source harnesses and skills for that purpose. Frontier models also support dynamic testing on a live website, but only after confirming you own the target, and they refuse targets you cannot prove. What changed 1. Per the author, open-source models, DeepSeek especially in its uncensored variant, are now good at finding vulnerabilities through dynamic testing, and they do not gate on target ownership. Presented as the author's assessment, not a benchmark. 2. Open-source dynamic pentest harnesses now exist, which was not the case before. 3. Combination: a refusal-free open model plus a ready-made harness removes the provider permission flow from the equation. The named harnesses The article analyzes three harnesses mentioned in media coverage of recent hacks: Artex and Cairn are Chinese-made, Strix is US-based. The text reviewed here ends before the per-harness deep dive, so concrete capabilities of each remain unconfirmed. Caveats The DeepSeek claims lack independent verification in the piece. Uncensored means refusal training removed, a policy change rather than added capability. Frontier ownership gates are likewise policy controls, so the meaningful change is procedural: dynamic agentic testing no longer depends on a provider signing off. Operational takeaway for defenders: treat external-facing assets as reachable by automated agents running commodity tooling, and weight raw unauthenticated exposure higher. 🔹 AISecurity #LLM #Pentesting #Harness #OpenSource 🔗 Source: https://utkusen.substack.com/p/how-llm-based-penetration-testing

Replies (0)

No replies.