What happened
Black Hat USA 2026 ran August 1 through 6 at the Mandalay Bay Convention Center in Las Vegas, with keynotes from White House National Cyber Director Sean Cairncross, Microsoft's David Weston on agentic security, and Arizona State University's Yan Shoshitaishvili on vulnerability research in the agentic age. The research theme, as Redmond Magazine put it, was AI accelerating familiar cyberattacks: faster exploitation of new vulnerabilities, more cloud-aware criminal activity, poisoned software dependencies, and identity attacks that turn legitimate logins into entry points. Two findings stood out for practical use. Researchers at Novee showed that a single malicious GitHub issue could steer AI coding agents from major vendors into remote code execution and credential theft. And 1Password's Off-By-1 Labs reported that 54 percent of AI-generated security patches failed to fix the original vulnerability. Conference coverage also highlighted the scale of machine identities, with figures cited of roughly 109 machine identities for every human one in enterprises.
Why it matters for your business
The takeaway for a small business is not that AI has invented new attacks. It is that the old ones now happen faster and at higher quality. Phishing emails read cleaner. New vulnerabilities get exploited in days, which shrinks the window between a patch existing and your systems needing it. And the research on coding agents matters to any business whose developers use AI tools: content from the outside world, like a GitHub issue or a pasted document, can carry instructions the AI will follow. The patch-quality finding cuts the other way and is just as useful. If someone on your team asks an AI to fix a security problem, more than half the time the fix may not actually work, so verification is not optional.
What to do about it
- Shorten your patch cycle where you can; monthly may no longer be fast enough for internet-facing systems.
- If your developers use AI coding agents, treat external content like issues and pull requests as untrusted input, and keep agents away from production credentials.
- Never ship an AI-generated security fix without a human review or a test that proves the hole is closed.