Black Hat 2026: what the keynotes argued

Black Hat USA ran from 1 to 6 August in Las Vegas, and for the first time the argument at the centre of the conference was not about a particular attack or defence. It was about economics.

David Weston, who leads agentic security at Microsoft, opened Wednesday with “The End of Rare: Defending When Offense Is Cheap.” His premise: the entire structure of modern security has been quietly priced around scarcity. Exploits are hard to write, so we assume network, process, identity and sandbox boundaries will mostly hold, and we invest in patching quickly and detecting known behaviour. That pricing is now wrong. Microsoft’s Security Response Center is processing nine times the vulnerability volume it handled in March, and it is doubling every six weeks. Weston’s line: hand-to-hand combat with attackers is a losing strategy for defence.

His evidence was uncomfortable. A Microsoft internal AI system took 200 Linux kernel vulnerabilities and turned 182 of them into crash-level proofs of concept — many working root exploits — at an average of $3.61 and 21 minutes each. Bugs he said used to take a year to craft by hand are being “spit out at industrial speed,” and not just on Windows.

The conclusion follows from the arithmetic. If finding bugs is nearly free and fixing them one at a time is not, the only remaining lever is to stop fixing them one at a time. Weston’s prescription: memory-safe languages so entire classes of bug cannot exist, formal verification so critical components are proven rather than tested, and automated remediation so the fixes still needed happen at the speed of discovery.

Yan Shoshitaishvili of Arizona State followed Thursday with “Vulnerability Research in the Agentic Age.” Where Weston spoke as a defender, Shoshitaishvili spoke as a researcher who has spent a career finding bugs. His account: agentic tools have shifted the human skill away from the finding itself and toward designing search strategies, validation pipelines and exploitability checks. He described a Linux kernel effort that went from hundreds of candidate findings to well over a thousand potential privilege escalations once the workflow was made aware of vulnerability properties — with the caveat that this was directional, not a controlled benchmark. He also flagged the problem we covered in Twenty-six thousand findings, two hundred fixes: discovery is outpacing anyone’s ability to handle the findings responsibly.

Of 121 briefings, 35 were directly about AI security, red teaming or LLM-assisted offence. The through-line, across keynotes and briefings: if bugs are cheap to find, the goal is to make them impossible rather than findable.