Black Hat 2026: Nobody Knows Who Your Robots Are
The biggest security conference of the year spent the week on one question. It’s a question you already have — you just have a smaller version of it.
A friend of mine is walking the Black Hat floor in Las Vegas this week. I’m in Virginia reading the wire feed with a second cup of coffee. Different vantage points, and I’d argue the wire feed is the better one this year — because the story isn’t sitting in any one booth. The story is that every booth is telling the same story.
That story goes: we handed AI agents the keys, and we never wrote down who we gave them to.
The industry stopped bolting on copilots
Two years ago the show floor was AI-as-feature. A chat box glued to a dashboard. This year it’s AI-as-governed-actor, and the vocabulary shifted with it. Zero Networks shipped something called Least Agency Enforcement — identity-based limits on what an agent can touch, what it can do, and when a human has to sign off. Sweet Security announced real-time blocking of rogue agents in live production. Cyera, Cycode, Rubrik, and a dozen others all landed some version of discover-the-agents, scope-the-agents, kill-the-agents-if-they-drift.
The clearest statement of the problem came from SailPoint on August 4, when they made their Agentic Fabric generally available and folded it together with the Identity Security Cloud under one banner. Buried in the announcement is the number that should stop you:
97% of AI agents have access to sensitive data. 21% of organizations are confident they can manage the risk.
That’s a 76-point gap between what we’ve deployed and what we understand. In my experience, gaps like that don’t close quietly.
Meanwhile the attacker side of the ledger got its own headline. CrowdStrike’s threat hunting report found AI now woven through adversary operations, with vulnerabilities exploited within hours of public disclosure. Cisco Talos published research built on recovered prompt logs — actual transcripts of attackers using language models to write malicious code and speed up vulnerability research. The theoretical phase of that conversation is over.
You already have non-human identities
Here’s where I want to be careful, because it would be easy to read all of this as enterprise noise. Five thousand endpoints, a SOC, a CISO, a budget with commas in it. Not your problem if you run a handful of WordPress sites.
Except the enterprise term for the thing they’re panicking about is “non-human identity,” and you have several. The WooCommerce API key. The service account your backup plugin authenticates with. The webhook secret between your form handler and your CRM. The GitHub personal access token sitting in a deploy script. The WordPress application password you generated in 2023 to test one thing and never revoked.
Every one of those is an identity that can act without a person present. Add an AI agent — a content generator with publish rights, an automation that opens tickets, an MCP server wired into your stack — and you’ve added actors that make decisions between your intent and your infrastructure. Same category. Smaller blast radius. Identical failure mode.
The good news is that “least agency” isn’t a product. It’s a design principle, and it costs nothing but an afternoon.
Open a spreadsheet. One row per key, token, service account, or app password that can touch your sites. Five columns: What it is. What it can do. When it expires. When you last rotated it. Whose name is on it.
Any row with a blank expiry, an unknown owner, or write access it doesn’t need is a finding. Fix those first. You will find at least one. I always do.
One more thing worth noting, since it landed the same week. The EU AI Act’s transparency obligations went live on August 2 — disclosure when someone’s talking to a machine, machine-readable marking of synthetic output, labeling for deepfakes. The heavier high-risk compliance rules got pushed to December 2027. If you build anything with an AI-facing surface for a client with European traffic, that’s a conversation to have now rather than in eighteen months.
I’m not writing this from the show floor. I’m writing it from a desk with a credentials list open in the other tab, and I’m three rows in and already uncomfortable. Which is, I think, roughly the point.
Sources
SailPoint, “SailPoint unified identity security for the AI era,” August 4, 2026 — sailpoint.com
SecurityWeek, “Black Hat USA 2026 – Summary of Vendor Announcements,” Parts 1–3 — securityweek.com
CSO Online, “The top new cybersecurity products at Black Hat USA 2026” — csoonline.com
EU AI Act transparency obligations, effective August 2, 2026.
