Spectra Assure Free Trial
Get your 14-day free trial of Spectra Assure for Software Supply Chain Security
Get Free TrialMore about Spectra Assure Free TrialIf Black Hat USA 2026 is any indication, the cybersecurity industry is going strong. This year's conference, part of a series of "Hacker Summer Camp" events held in Las Vegas during the first week of August, drew more than 23,000 verified attendees, according to organizer Informa Tech. That was an increase of more than 15% from last year’s show, with AI the dominant theme.
Case in point: a packed, last-minute Breaking News session called “Frontier AI Security Breach Exposed” in which OpenAI security engineers walked through how one of the company’s pre-release research models exploited a zero-day flaw to escape its testing sandbox, reach the open internet, and compromise AI model repository Hugging Face.
The OpenAI talk got lots of coverage, but it was just one of a series of top tier talks focused on AI threats. Conference keynotes featured a quartet of federal cyber officials (Sean Cairncross, the White House’s national cyber director; Nick Andersen, acting director of the U.S. Cybersecurity and Infrastructure Security Agency (CISA); Katherine E. Sutton, assistant secretary of war for cyber policy; and Brett Leatherman, assistant director of the FBI’s Cyber Division).
All talks emphasized that AI-driven threats have become central to national security. The urgency of their message was underscored by the fact that CISA and other agencies recently steered clear of Black Hat, DEF CON and other cyber industry events.
The conference revealed an information security industry in the midst of a tectonic shift, as large language models (LLMs) and agentic AI transform both cyber threats and protections.
Here are three key takeaways that stood out at this year’s Black Hat.
[ See webinar: Autonomy, Not Autopilot: Get Real About the Agentic SOC ]
OpenAI’s presentation on the Hugging Face breach demonstrated how fast agentic AI risks are materializing. OpenAI’s Michael Dalton described how an unreleased model, during an evaluation that began in May, hit a roadblock, reasoned its way into writing files on Artifactory, a connected third-party repository, and left notes there that other AI agents found and built on — organizing themselves through a message board they created. The agents uncovered a remote code execution flaw and an admin-privilege bug. After OpenAI patched the initial issues, the agents regrouped and pushed through to Hugging Face.
“In the near future, we should expect that threat actors will intentionally deploy, optimize, weaponize, and use offensive agent collectives in the manner that we have just described here.”
—Michael Dalton
And that incident wasn’t a one-off. Earlier this month, the United Kingdom’s AI Security Institute issued a report saying that during permissive testing, a model attempted a supply chain attack on a real open-source project. The attack, which involved fake identities and social engineering aimed at human maintainers, was the first time AISI had observed autonomy and deception risks manifest in the wild, unprompted.
LLM-powered tools are rewriting the rules of the security industry. They can scan software and binaries for known and zero-day flaws at machine speed then use their acute reasoning to chain together long strings of minor, moderate and critical weaknesses to seize control of a target system. They thus become super adversaries that leave the find-and-patch model in the dust by assembling working exploit chains from issues nobody flagged as urgent.
Countering such threats requires more than faster scanning. Organizations need deep insight into the context of their own environment — their assets, dependencies, and exposure — and then deploy defensive AI that can spot the kind of customized, local compromise that generic threat feeds miss. To maximize the effectiveness of the defensive AI, they need to exploit its ability to identify discrete patterns — the more data the better, said Eric Thoen, ReversingLabs’ vice president of product management.
“It’s garbage in, garbage out. If you don’t have good, high-quality deterministic context, you’re probably not going to get the right answers from your AI.”
—Eric Thoen.
What once amounted to cybersecurity noise that had to be discarded in favor of high-fidelity indicators can now drive more accurate and more cost-efficient AI-based detection and resolution. This AI-vs.-AI world will remake the cybersecurity market.
It is very clear now that security teams can’t fight machine-speed adversaries with human-speed tools and operations. AI-powered threats assess targets, identify flaws, tailor attacks, and adapt to changing conditions faster than any human team can ever hope to match. It was that fact that led to the creation of the Agentic SOC Alliance, which ExtraHop and 14 other companies, including ReversingLabs, introduced in July.
The alliance has proposed an open architecture built on three layers: context, in the form of an operational knowledge graph; a governed AI runtime; and an interchangeable reasoning model. Together, the layers are designed to let autonomous agents detect, decide, and respond at machine speed.
“Post-Mythos AI has fundamentally changed cyber defense,” ExtraHop CEO Greg Clark said, referring to the frontier AI model of Anthropic’s Claude.
“The industry needs a blueprint for autonomous security.”
—Greg Clark
Fiserv CISO Jason Dewez put it more bluntly: The traditional SIEM model, built for human analysts working at human speed, will not keep up.
Today, security teams within organizations are overwhelmed by alerts, and ReversingLabs CEO Mario Vuksan noted that they have always struggled to determine how many — and which — of those they really need to inspect.
Today, with agentic AI powering attacks, the answer is simple: “Every single one of them,” Vuksan said. To do that, the industry now has to figure out which technologies are needed and how they can be tailored to optimize AI-powered threat detection and mitigation. The Agentic SOC Alliance was created to help answer that question, Vuksan said.
All of these changes have happened fast, in just months, and the months ahead will test how fast the industry can adapt. AI-powered threats aren’t slowing down, and the incidents detailed at Black Hat suggest that they’re outpacing conventional defenses. At a private event during the show, ExtraHop and other members of the Agentic SOC Alliance appealed to fellow security firms to join and bring their knowledge and expertise to the Alliance.
In the wake of this year’s Black Hat, we can expect more organizations to go beyond bolting AI onto existing tools and fundamentally rethinking their security operations and technology stack. And those that pair rich, environment-specific context with AI-powered security tooling will fare best.
Change is a constant-- in cybersecurity even more than in other pursuits. The last three decades saw it evolve from simple antivirus and firewall deployments to fully equipped SOCs staffed by security operations teams working complex kill chains to counter a wide range of threats. But the industry is about to be buffeted by the biggest change of all from an influx of AI-powered threats and defenses.
“[This] will change the security processes and introduce new positions, new titles, and new ways of addressing risks that will be with us forever.”
—Mario Vuksan
[ See webinar: Why Binary Analysis Has Become the Standard for Software Risk ]
Researchers built a worm that reasons about hosts it infects, and the open-weight models powering it sit outside AI-provider safety controls.
While prompt injection and data disclosure remain concerns, excessive agency climbed the list — not surprising with recent security incidents.
The post-mortems of two compromises by rogue AI agents show that security teams need to focus on guardrails, not the AI model.