A TheGRCGuy Recap
Every August, the security industry decamps to Las Vegas for what's affectionately known as "Hacker Summer Camp" — Black Hat USA, DEF CON, and BSides Las Vegas, back to back to back. This year's run (Black Hat: August 1–6 at Mandalay Bay; DEF CON 34: August 6–9 at the Las Vegas Convention Center West Hall) drew over 23,000 verified Black Hat attendees alone, more than 15% growth year over year, plus an estimated 30,000 at DEF CON. If you weren't on the ground, here's the GRC-relevant signal from the noise.
The Headline Theme: Agentic AI Is Now a Governance Problem, Not Just a Security One
If 2025 was the year organizations figured out how to build AI agents, 2026 was the year everyone had to figure out how to govern them. That shift showed up everywhere — in keynotes, in vendor pitches, and in the DEF CON badge theme itself ("Agency"). A few data points worth pulling into your risk register conversations:
Machine identities are outpacing human ones, fast. Palo Alto Networks reported a 109:1 machine-to-human identity ratio, up from 82:1 last year. Separate research from Cyera/Oasis Security put the non-human identity ratio at 45:1, with 78% of organizations lacking formal policies for managing AI identities at all.
AI-agent incidents are already common, not hypothetical. Research cited at the show found 88% of organizations had experienced a confirmed or suspected AI-agent security incident.
Exhibitor behavior tells the story. An independent floor analysis found that of roughly 450 Black Hat exhibitors, 235 were marketing AI or agentic capabilities — over half the Business Hall.
The practical takeaway for GRC teams: AI agent inventory and lifecycle governance can no longer sit as a future-state item in your roadmap. If your organization has agents touching production systems, code repositories, or customer data, that's an asset class your risk register and access-governance program need to account for now, with the same rigor you'd apply to service accounts or third-party integrations.
"The Harness Wars": Vendors Are Racing to Wrap AI in Guardrails
Forrester's recap coined a memorable phrase for a trend that ran through nearly every keynote: the rise of the "harness" — the layer of tooling meant to constrain, monitor, and validate what AI agents and models are allowed to do. Microsoft's David Weston argued in his keynote that AI is fundamentally changing attacker economics: as vulnerability discovery and exploitation get cheaper thanks to AI, "patch faster" and "detect and respond" become less reliable as standalone strategies.
That framing matters for how we think about control design. If AI compresses the time between vulnerability discovery and exploitation, GRC programs built around periodic control testing and quarterly vulnerability scans are going to feel increasingly out of step with the actual threat tempo. Expect more pressure toward continuous control monitoring and real-time exposure validation as a baseline expectation, not a maturity-model aspiration.
The Incident That Made It Real: ChainDrop
Timing gave this year's conferences an unusually sharp illustration of the agentic risk conversation. During Black Hat week, Microsoft disclosed ChainDrop, a self-propagating npm worm that specifically targeted configuration files for AI coding assistants (including Claude Code and VS Code) to hijack them as persistence vectors. DEF CON's AI Village had flagged agentic worm risk as a coming threat just days before ChainDrop hit — a reminder that "AI supply chain security" isn't an abstract category anymore.
Related research at DEF CON — dubbed "agentjacking" — showed how something as mundane as a publicly exposed error-tracking credential, chained with an MCP (Model Context Protocol) integration, could be escalated into remote code execution. Worth flagging to your engineering and DevSecOps counterparts if your organization has adopted MCP-based tooling: credential hygiene around these integrations deserves a fresh look.
Identity Is Still the Story Underneath the AI Story
Away from the AI headlines, identity compromise remained the quiet constant. SpyCloud marked its tenth year of tracking stolen credentials by surpassing one trillion recaptured identity assets, and reported that nearly half of what it now recovers is stolen session cookies and tokens rather than passwords — a sign that credential theft has moved past the password entirely. Its 2026 Identity Exposure Report also flagged 18.1 million exposed API keys and tokens, plus 6.2 million credentials tied specifically to AI tools.
Separately, research on passkey and passwordless implementations found that adoption alone doesn't eliminate risk — account recovery flows remain a soft spot that even low-skilled attackers can exploit. For anyone advising on identity-related controls (NIST 800-53 IA family, ISO 27001 Annex A identity management controls, or CCPA/CPRA authentication requirements), this is a good prompt to make sure recovery-flow testing is explicitly in scope, not just primary authentication.
DEF CON's AI Village Formalized Autonomous Offense
DEF CON leaned hard into its "Agency" theme with the debut of HALctf, described as DEF CON's first fully autonomous-only Capture the Flag competition — participants deploy self-contained AI agents built on open-source models and let them scout, exploit, and pivot through live targets with no human intervention. It's a natural next step after a 2025 incident where a CTF team's autonomous agent reportedly solved and won a challenge on its own, without the team realizing it until a judge flagged it.
Alongside HALctf, DEF CON ran more than 40 other competitions spanning the usual breadth of the conference's villages: Capture the Packet, Car Hacking Village CTF, the Aerospace Village's STARPWN, the Biohacking Village's medical-device-focused Code Cadaver, and the long-running Crack Me If You Can password-cracking marathon.
Policy and Legal Threads Worth Tracking
A few sessions with direct compliance and public-policy relevance:
A retrospective from DOJ's Leonard Bailey traced the multi-year, practitioner-driven effort to reform U.S. computer hacking law from within the department — a useful reminder that the legal environment security researchers operate under has been actively shifting.
EFF researchers highlighted growing global momentum behind age-verification and online-safety legislation, arguing that laws framed around protecting children are, in practice, introducing new data breach and privacy-misuse risk by requiring broader identity verification infrastructure. Worth watching if your organization operates in jurisdictions considering similar rules.
DEF CON, along with the U.S. Air Force, several courts, and other venues, moved to restrict Meta-style recording-capable smart glasses over surveillance concerns — a small but telling sign of how physical privacy policy is starting to catch up with wearable AI hardware.
What This Means for Your Program
Pulling it together, a few concrete actions worth considering coming out of this year's conferences:
Stand up (or mature) an AI/agent asset inventory. You can't govern what you haven't catalogued. Agents, their permissions, and the systems they can touch need a home in your existing asset and access management processes.
Revisit vulnerability response SLAs against an AI-accelerated threat timeline. If exploitation is getting cheaper and faster, patch cadence and detection thresholds calibrated to human-speed attacks may already be lagging.
Extend identity and credential controls explicitly to AI tooling. MCP integrations, coding assistants, and agent service accounts need the same credential hygiene and least-privilege review as any other system integration — arguably more, given how new the tooling is.
Test recovery flows, not just primary authentication, as passwordless and passkey adoption grows.
Watch the age-verification and biometric legislation space if you operate in consumer-facing or multi-jurisdictional environments — this is shaping up to be a live regulatory front.
Hacker Summer Camp has a way of previewing, a year or two early, what shows up in frameworks, audits, and board questions later. This year's clearest signal: agentic AI has moved from an innovation topic to a governance obligation. The organizations that get ahead of that shift now will have a much easier 2027 audit cycle than the ones that treat it as next year's problem.
Sources: Forrester, TechTarget/Dark Reading, SecurityWeek, CSO Online, Help Net Security, SpyCloud, Black Hat/Informa, DEF CON, and TechTimes coverage of Black Hat USA 2026 and DEF CON 34.