- A speaker at Black Hat USA 2026 said defenders should use frontier AI to red-team their own systems continuously, because attackers will use open-source models.
- Morgan Adamski of PwC warned against assuming that more compliance requirements will make a business more secure, and urged firms to build agility.
At a Black Hat USA 2026 panel on AI and cyber defense, one speaker urged defenders to attack their own systems with frontier AI models, and to keep doing it. The speaker said attackers will use open-source models for the same job. Morgan Adamski of PwC warned against assuming that more compliance requirements will make a firm more secure. The transcript does not name the first speaker, so we call them a speaker on the show.
What was said
In the closing round, the host asked each panelist for one piece of advice. That speaker listed three things: inventory your agents, create audit trails, and use frontier models for aggressive red teaming. Red teaming means testing your defenses the way an attacker would.
The speaker also wanted continuous testing of controls, to catch regressions. A regression is a safeguard that once worked and has quietly stopped. The reason given was short: "attackers will use open source models to do that." The speaker added that requiring a human to approve every defensive action will not scale. People should still be responsible for policy.
Adamski took a different angle. She said: "Do not assume that just applying a significant amount of compliance requirements and oversight on things, it's going to somehow make it better and more secure." She told firms to look at what makes sense for their own business. She also asked whether they have built enough agility for the unexpected.
Why it matters
Our reading: a yearly penetration test or a passed audit is a snapshot. If attackers can probe at any hour, a snapshot ages fast. Managers can ask who tests their defenses, how often, and whether the tests rerun after every change.
Here is WebPulse's plain explanation of continuous testing. The panel did not describe it this way. A model tries moves an attacker might try against your systems, and the results are saved. After each change, the test runs again and is compared with the last run. If a control that caught something before now misses it, that is a regression. One example is a detection rule that stops firing after an update.
Agents here means AI agents, which are programs that act on their own. An audit trail is a record of what each one did. Fotis Chantzis of OpenAI said you cannot really authorize what you cannot identify, and that includes agents. That fits the advice to inventory them.
The other side
The panel left real gaps. Nobody in these excerpts said open-source models are free or as capable as frontier ones. Calling defenders who skip this testing second is our framing, not the panel's. No one offered evidence that continuous AI red teaming prevents breaches, or said what it costs.
Finding flaws is also not fixing them. Katie Moussouris said the industry could never patch its way out of this. Adamski said there may not be a patch for every vulnerability that AI discovers. More findings could mean a longer list of known, unfixed problems.
Adamski did not reject compliance outright. She asked firms to pick the requirements that fit their business. Two panelists also work for frontier AI developers, OpenAI and Anthropic, which build the kind of tools being recommended.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- Black Hat: Black Hat USA 2026 | AI and the Future of Cyber Defense Panel (2026-11-13)





