The industry-wide reaction to Claude Fable 5 and its research-tier counterpart Mythos did not surprise me. It felt like a long time coming. Anyone using AI for software development and security has watched these systems move from being productivity assistants to something more like an active testing partner.
In our own security work, AI has already changed the math. In one internal experiment, an AI testing workflow reproduced the findings of a traditional penetration test that had cost us $20,000 and taken two weeks. It did so in under two days, for a few hundred dollars in compute. It was not a perfect replacement for human expertise, but it showed how discovery is becoming cheap, fast, and always available.
Frontier AI is pushing the industry toward greater autonomy. Models are moving from reconnaissance to exploitation to post-exploitation in one connected flow, with less human feedback required along the way. Current frontier models can already reproduce known flaws and generate proof-of-concept paths very quickly. This next phase pushes the trend toward full-chain automation.
And yes, these are considerable breakthroughs in discovery. Finding more issues is useful, but the industry would be remiss to stop there. The tougher part starts immediately after discovery. Someone still has to decide whether a threat is real and what downstream effects it could have on the business. Teams then need to determine who owns the fix, how to test the patch, and how to prove the business is safer when the ticket closes.
See also: Domain-Specific LLMs: How to Make AI Useful for Your Business
Speed Up Patching Without Losing Assurance
Remediation is where security work slows down. Patching involves stakeholders, feature teams, release windows, tests, rollback plans, and the fear of breaking software users depend on. Those checks exist for good reasons. The issue is speed. AI will increase the number of findings, while many organizations still patch manually.
Security teams need technical controls to validate patches at a faster pace. A strong remediation workflow should answer basic questions automatically where possible. Teams need to know whether the patch build passes regression tests and if the exploit path still works. They also need visibility into whether the affected code reached production, whether a new dependency introduced additional risk, and whether the change can be rolled back safely.
That requires tighter coordination across vulnerability management and CI/CD workflows. Test automation, service ownership, and runtime inventory also need to be more closely connected.
Assume a Breach to Reduce Blast Radius and Govern Agents
As discovery gets cheaper, the conversation cannot stop at reducing vulnerability backlogs. If attack automation becomes faster and more autonomous, defenders must improve how they respond to a breach. An assume-breach mindset prepares security teams for that reality.
In practice, this means prioritizing detection, containment, and recovery alongside prevention. Organizations need to be able to detect credential misuse immediately and revoke or rotate access before an attacker can move laterally. Alerts also need to be clear enough to trigger action. AI-driven attacks compress the time between compromise and exploitation, which will expose organizations with slow detection and response processes.
Reducing the blast radius is one of the most practical results of an assume-breach mindset. Once organizations accept that some attacks will succeed, the focus shifts toward limiting how much damage can happen afterward. Systems become easier to contain, access becomes more deliberate, and a single compromise is less likely to spread across the environment.
The same thinking applies to AI agents. Solid governance depends on placing clear boundaries around what agents are allowed to do and where human oversight is still necessary. Organizations that do not build with assume-breach in mind tend to accumulate overly broad access and loosely controlled automation. That may not create immediate problems, but it increases the likelihood that one mistake or compromise turns into something bigger.
Use AI Before Attackers Do
Waiting for a perfect AI security tool is the wrong posture. Organizations should use the models available now to test their systems. The goal is controlled pressure testing to determine whether a model can identify exposed paths, weak configurations, unsafe dependencies, credential risks, or poor assumptions before a real attacker does.
In this process, it’s important to understand that, while AI can identify potential vulnerabilities and even produce proof-of-concept exploits, the findings still require human validation. Security teams must determine whether the target is actually reachable and what privileges or authentication are involved. From there, they can assess what data or credentials may be exposed, which controls are in place, and how critical a given asset is to the business.
Having the longest list of findings is not the goal. It’s far more important to have reliable context. Teams need clear visibility into their riskiest applications and vulnerabilities, and how those connect to overall security posture.
Build the Operating Model Now
I think in the next year the fundamentals of security will run at a higher speed. Organizations will still need strong asset ownership and credential hygiene, along with patch validation, detection, and response capabilities. The difference is the pace at which weak points can be found and chained together.
Organizations should use accessible frontier AI models now, inside safe boundaries, to test their security. Inventory critical applications and systems and vet vulnerabilities against business context. Automate patch validation whenever possible and build agent guardrails before AI systems ever touch production.
Frontier AI is here, and the industry must adapt to the changes it brings. Security teams can still get ahead by testing and protecting their systems before attackers discover widespread vulnerabilities and novel ways to exploit them.
