COMMENTARY: Anthropic’s Claude Mythos Preview has landed with the kind of reception security leaders can’t dismiss. Anthropic says the model has already identified thousands of high-severity vulnerabilities, including in major operating systems and web browsers, and has limited access to it through Project Glasswing rather than making it broadly available.Reuters has also reported that regulators and major financial institutions are assessing the implications. All of these recent developments does not mean CISOs should torch their roadmaps and stand up an entirely new AI security program.[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]But it does mean the economics of offense has shifted again. If frontier models like Claude Mythos can shorten the time it takes attackers to scout, build exploits, and persuade targets, then any delay by defenders makes weaknesses more dangerous. The fundamentals still matter, but the tempo changes.My advice: Don’t rebuild the program. Reinforce the parts that carry the most weight.Just don’t panic. And don’t remain passive. CISOs definitely don’t need a revamp for the sake of a checklist, which simply renames existing workstreams and burns budget on new labels. They need to move a handful of controls higher on the priority stack and make them harder to bypass. Let’s look at six actions that stand out:Claude Mythos should not trigger a ceremonial overhaul of the security program, and it’s definitely not a reason to wait for the next model announcement. We have to find a middle ground: keep the architecture, but raise the urgency. Security leaders do not need a new doctrine. They need tighter identity controls, faster exposure reduction, better governance over AI touchpoints, and more realistic testing.David Balaban, owner, Privacy-PCSC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.
- Treat AI exposure as an attack surface: Most enterprises still do not have a clean view of where LLMs touch sensitive workflows. That blind spot has now become a security problem. Map where employees, developers, copilots, agents and vendors can feed models internal data or let models act on enterprise systems. If a model can read a ticket, summarize a document, write code, call an API or influence a decision, it belongs in threat modeling, data classification and logging.
- Re-rank identity controls before phishing gets better again: Security teams have spent years talking about identity as the new perimeter. Claude Mythos makes that cliché more operational. The concern is not only better phishing emails: it’s better pretexting, better impersonation, better lures tailored to role, project and timing. That argues for phishing-resistant MFA, stricter privileged access reviews, stronger help-desk verification, and tighter approval paths for password resets, wire changes and admin actions that still depend too heavily on trust.
- Shorten the window between exposure and mitigation: If capable models reduce the time required to analyze public code, weaponize an N-day, or chain together misconfigurations, then long patch cycles become even more expensive. CISOs should focus less on broad promises about “faster patching” and more on concrete bottlenecks: asset visibility, ownership, maintenance windows, exception handling and compensating controls. We don’t need perfection. But we do need to reduce the number of exploitable issues that remain both known and reachable for too long.
- Put guardrails around AI-assisted development: Many organizations are rightly excited about developer productivity. They should be equally interested in how fast insecure code, leaked secrets and fragile dependencies can now move through the pipeline. Secure coding programs need to explicitly account for AI-generated output through mandatory secrets scanning, stricter code review for sensitive functions, provenance checks for generated packages, and approval gates for agents that can write or execute changes. It’s not about whether developers use AI, but whether security can verify what AI helps ship.
- Run exercises that assume the attacker also has automation: Tabletop exercises built around yesterday’s attack pace will not tell CISOs much about tomorrow’s resilience. Purple-team scenarios should test what happens when phishing volume spikes, exploit chaining accelerates, or recon against public assets becomes far more precise. Security leaders should ask a simple question: where does the organization still depend on the attacker being slow, noisy or under-resourced? That’s where a Claude Mythos-class capability will most likely expose brittle assumptions.
- Push third parties harder on assurance: Vendors will increasingly market AI features as differentiators, often before governance catches up. CISOs should require clearer answers about model access, data handling, retention, prompt logging, human review, agent permissions and kill switches. Just as important, they should translate this issue for the board in operational terms: which risks are genuinely new, which are familiar but accelerated, and which metrics will show whether the company will get more secure or merely more fashionable.
