Blog & Insights
From the lab to the field — deep analysis on AI Security, Offensive Security and the latest trends from the DNA team
Project Glasswing: What Mythos Showed Us
In recent weeks, Cloudflare pointed Mythos Preview and other security-focused LLMs at live code across critical parts of their infrastructure. They share what they observed, the models' strengths and weaknesses, and what the work around them needs to look like before any of it can scale.
Read moreAfter Automation
AI progress creates more work for humans, not less. Dan Shipper, CEO of Every, walks through why the more they automate at Every, the more expert human work appears — and why that pattern doesn't break even when we reach AGI.
The Downfall of Bug Bounties
Shubham Shah on why bug bounty platforms have stopped delivering for serious researchers — AI slop reports flooding triage queues, painful delays on critical bugs, and platforms that no longer understand the hacker community they were built to serve.
GPT-5.5: Mythos-Like Hacking, Open To All
Anthropic has Mythos, but only a select few have seen it. Now OpenAI has GPT-5.5 — a comparable step change in vulnerability detection — released freely. XBOW had early access and benchmarked it against real pentesting workflows. Here's what they found.
AI for Pentesting: Strengths, Weaknesses, and Where XBOW Fills the Gaps
AI is transforming pentesting, but there are areas where it's strong, like pattern matching, and areas where it is weak, like orchestration. XBOW leverages the strengths and adds scaffolding to the weaknesses to create an enterprise-ready autonomous offensive security platform.
How We Hacked McKinsey's AI Platform
McKinsey's internal AI platform Lilli — used by 43,000+ employees — was compromised by an autonomous offensive agent in under 2 hours. No credentials. No insider knowledge. 46.5 million chat messages exposed.
Emergent Cyber Behavior: When AI Agents Become Offensive Threat Actors
AI agents deployed for routine enterprise tasks are autonomously hacking the systems they operate in. No one asked them to. No adversarial prompting was involved.