Get ready for Black Hat USA 2026 

Open source malware is no longer rare: 51% of organizations found malicious packages in 2025. Yet the gap is striking — 88% know new releases are the riskiest window, but only 21% enforce a cooldown. New data from 605 security professionals shows where defenses are falling short.

Agents are powerful tools for coding tasks, but many teams still use them as blunt instruments for security tasks, burning time and tokens. Agents shouldn't have to reconstruct your security posture from scratch on every task; they need direct access to vulnerability details, upgrade paths, package impact analysis, and remediation guidance specific to your codebase. This is the role of harness engineering. We'll show how building specialized agents cut cost, improve outcomes, and give you visibility into what your agents are actually doing.

Think you know AI pentesting? Put it to the test at this fast-paced, interactive trivia session. AI Pentesting Trivia Showdown challenges participants with questions spanning offensive security fundamentals, real-world attack paths, AI-assisted testing concepts, vulnerability validation, and modern application security practices. 

Designed for practitioners of all experience levels, the game blends learning with competition in a format that is approachable, engaging, and fun to watch or join. Attendees will sharpen their understanding of how AI is changing penetration testing, pick up practical security insights, and leave with a stronger grasp of modern offensive techniques and terminology.

This session will showcase a learning lab walking through the XBOW/AWS integration - how XBOW discovers, enumerates, and tests an example customer's AWS org end to end, from asset discovery to validated findings.

Pointed at buffa, Anthropic's Rust protobuf library, Endor Labs flagged a data flow that's easy to skim past: an unknown-field decoder that allocates heap in proportion to attacker-controlled input. The obvious sink amplifies a modest ~2x, but one branch further sits a second sink that turns a tiny input into a ~22x allocation, enough to OOM-kill a process (now CVE-2026-55407). A frontier model reasons well about the code it surfaces but leaves most of the repo unexamined, while a pattern scanner matches everywhere but is blind to behavior; we'll walk through how following the data flow catches what both miss.