Artificial intelligence (AI) and large language models (LLMs) have dominated headlines, boardroom discussions, and company roadmaps since late 2022. Even so, there have been watershed moments that changed the nature of the discussion - for example, the release of Anthropic’s Claude Code and Opus 4.6 model demonstrated a step change in LLM coding capabilities, forcing engineering teams to readjust timelines and priorities. 

In recent months, cybersecurity is experiencing its own ‘Opus 4.6 moment’, with early news of hypercapable models that can find thousands of new vulnerabilities in existing software, at a tiny fraction of the cost of traditional threat research. In this paper, we will look at the current state of AI-powered exposure management, explain how this shifts the balance between attackers and defenders, and offer practical suggestions for security organizations.

Claude Mythos, Fable, and the Coming Wave of Hypercapable AI Models for Cybersecurity

In April 2026, Anthropic announced Claude Mythos Preview, an unreleased general-purpose model with unusually strong cybersecurity abilities. Citing misuse risk, Anthropic opted not to ship the model publicly; instead, it has initially gated access to roughly fifty partner organizations through Project Glasswing. 

A short while later, Anthropic reported that partners had used the model to find more than 10,000 high- or critical-severity vulnerabilities across mission-critical software. A parallel scan of more than 1,000 open-source projects flagged 23,019 candidate issues, 6,202 of them rated high or critical; of a sampled subset reviewed by outside firms, 90.6% were confirmed as valid. Headline examples include CVE-2026-5194, a critical certificate-forgery flaw in wolfSSL (embedded in an estimated five billion devices); as well as a 27-year-old denial-of-service bug in the OpenBSD security OS, which would enable an attacker to remotely crash any machine running it - firewalls, server environments, and more.

By early June, Anthropic had shipped a version of Mythos to the public, along with Fable 5 - a model from the same family. Yet within days, an unexpected U.S. government directive effectively banned the models, citing national security concerns. As of publication, it remains to be seen whether Anthropic will find a way to re-release these models. 

Since few researchers outside of Anthropic have had a chance to thoroughly test Mythos and Fable, some open questions remain about these models’ actual capabilities and uniqueness. But regardless of specific, debatable claims, the direction of travel is clear: hypercapable models are coming for cybersecurity. Where does this leave the industry?

AI Favors Defenders in the Long Term, but it can Overwhelm Them in the Short Term

New AI models will change the economics, as well as the mechanics, of exposure management by shifting the equilibrium between attackers and defenders. To understand this concept, we need to first look at the current situation, and why it has traditionally tended toward stability over time.

The security equilibrium

Cybersecurity is always an arms race between defenders and attackers, but they have a different set of challenges to overcome:

• Attackers only need to find one vulnerability, but they need to exploit it in great depth to cause actual damage - that is, to turn a theoretically vulnerable code path into a reliable, repeatable exploit that defeats mitigations (ASLR, signed pointers, stack canaries, sandboxing…) and works consistently across the target's real versions and configurations. Exploitable weaknesses are hard and expensive to find, and even harder to operationalize; vulnerability research is often the most costly part of offensive cyber.
• Defenders need to cover an ever-expanding attack surface - but they just need to block the attack path, even in a shallow way. You do not need to understand the full exploit trajectory to prevent its abuse in the wild. This has allowed defenders to use more automation and software, where investment has grown over the past decades.

In the past, technology has accelerated the capabilities of both sides, so that the overall state of security remained more or less steady. For example, cloud migration created new surfaces to attack, and new exposure points that diligent attackers could find; but it created new defensive monitoring and scanning tools (e.g., by reducing dependence on local agents), which were useful for defenders.

So are we going to see more of the same with LLMs? In our opinion, not necessarily.

AI shifts the bottleneck from finding to fixing

Previous technological shifts created degrees of automation for low-value work - allowing threat researchers to expend less effort on tedious busywork such as setting up fuzzing harnesses and triaging the resulting crash dumps, and more on finding the ‘juicy’ vulnerabilities that were ripe for exploitation. Hypercapable AI is different in that it promises to automate the highest-value, most expensive part of the work - discovering exploitable weaknesses.

If an AI model can discover critical vulnerabilities in a few hours or days of work - even if the token price reaches thousands of dollars - it has deeply changed the economics of exposure management. It dramatically compresses the effort required to weaponize complex bugs, and reduces the associated price tag to a small fraction of what it previously was (since human expertise was the largest ‘line item’). Analyzing very large codebases and stacks for subtle issues will become a run-of-the-mill capability. On the defender side, it shifts the focus to verification, prioritization, and remediation.

Long-term opportunities, short-term threats

We believe that over the medium-to-long term, this shift will favor the defenders, and make software safer than it is today. There are several reasons for this conclusion:

• As we’ve mentioned above, blocking an attack is significantly easier than perpetrating one. If both sides have access to the same models and can detect the same vulnerabilities, defenders have the advantage of being able to plug these weaknesses faster than attackers can exploit them.
• Defenders have white-box access: they can aim these models at their own source, build systems, and telemetry, while attackers work black-box from the outside - which makes for more powerful AI discovery. For software they build and operate, they also control the fix, so they can close the gap before the same finding is weaponized.
• Much of the current rise in CVEs is due to vulnerabilities detected in existing software and open-source tools. As AI becomes embedded earlier in the development lifecycle, it will help teams identify these issues software is released.

However, before this optimistic vision pans out, security teams will face an uphill battle. Things will get worse before they get better. In the short term, AI-powered exposure detection puts defenders at a significant disadvantage.

As new and more capable AI models are deployed at a wider scale, they are likely to open the floodgates - exposing thousands of vulnerabilities in the software that organizations are already running. According to Anthropic’s announcements, less than 1% of the vulnerabilities discovered by Mythos have been patched so far, and the backlog will accumulate even faster with better models and wider usage. 

This will create immense pressure to patch, fix, and replace legacy tools before newly discovered vulnerabilities are exploited in the wild. Security teams are already buckling under the pressure, understaffed and underfunded; this massive new workload could overwhelm them, leaving organizations exposed.

How to Win the Race Against AI-powered Exposure Detection

New cyber threats can be likened to new biological viruses. Organizations, like human bodies, develop immunity over time; with AI, we are at the early stages, and must now consider the best ways to build that immunity more quickly.

The way to win this race is to stop treating discovery as the core problem. With AI the real bottleneck is speed: how quickly defenders can decide what matters, and how quickly they can fix it without breaking the business.

Prioritization cannot start with CVSS alone. It has to start with the threat model. An opportunistic attacker harvesting credentials through malicious packages, a ransomware operator looking for the fastest path to domain control, and a state-backed actor targeting SCADA or edge infrastructure do not create the same risk, and should not drive the same remediation queue.

Once the threat is understood, defenders can identify the crown jewels, map the realistic attack paths into them, and focus on the vulnerabilities that are actually reachable, exploitable, and impactful in their environment. 

From there, remediation has to become both more automated and more intelligent. Not just faster ticketing, but faster exposure reduction. That means investing in the changes that remove the most risk with the least operational friction: resilient patching paths, safer upgrade mechanisms, redundancy, reversible migrations, separated data layers, load-balanced traffic shifting, and platforms that support safe in-place updates.

It also means using AI defensively for triage, enrichment, ticket routing, validation, and low-risk remediation actions under clear guardrails. And when patching is too slow, compensating controls such as segmentation, configuration changes, access restrictions, and traffic filtering should be treated as first-class remediation paths, not second-best substitutes.

In the AI era, exposure management becomes a race between automated discovery and operational remediation. The winners will be the defenders who can turn context into action before attackers turn findings into exploits - and this will only be achieved by combining threat-driven prioritization and intelligent automated response.Industry Perspectives: Four Action Items to Prioritize

 We asked domain experts from Team8 portfolio companies to offer one piece of advice on what security leaders should do next.

Automate exposure operations

Dannielle Shanahan-Rios, Field Marketing Managerat Nagomi Security

To keep pace with AI-powered attackers, companies must have autonomous workflows that go beyond identifying threats to neutralizing them. This can be achieved through Agentic Exposure Operations - an approach which proactively identifies exposures and toxic combinations based on organizations' unique environments (misconfigurations, missing controls, and other contextual factors) rather than generic signals. This higher-granularity view of the attack surface enables SecOps teams to prioritize fixes appropriately by deploying controls, hardening configurations, patching, or even blocking when necessary. 

Prepare for Shrinking SLAs

Ilya Kleyman, CMO at Ionix

The traditional SLA window no longer closes fast enough to keep attackers out. Haphazard, cumbersome patching across an opaque attack surface is doomed to failure because every single exposure will be discovered and exploited at lightning speed. To survive, firms will need comprehensive visibility, pin-point validation, and agentic mitigation. The good news is that with proper real-time discovery and validation in place, machine speed defense is indeed possible. With that in mind, Ionix recently committed to a 12-hour SLA for live exposure defense. 

Focus on reachability

Alon Shamir, Head of Marketing at Astelia

With new models, severity scores stop telling you much because nearly every finding starts to look exploitable. Reachability analysis gives organizations a way to regain that footing. It asks whether an attacker can actually reach a vulnerability and lets the network itself answer, checking each finding for the path, privileges, and conditions a real exploit would need. Organizations working this way can cut an overflowing backlog of "criticals" down to the handful an attacker can abuse - and initiate multiple remediation paths to close those routes quickly: segmentation, configuration changes, or compensating controls. 

Build attacker-like validation

Noa Hen, CPO at Alt

The current pentesting and vulnerability-management model is no longer fit for purpose when hundreds of releases and thousands of changes are completed between annual pentests. Point-in-time assurance no longer reflects risk when attackers use AI to find full attack paths rather than isolated findings. Instead, CISOs should build continuous, attacker-like validation into their program. In practice, this means deploying AI “attackers” on your side before real attackers arrive - using agentic pentesting and continuous red-teaming to expand coverage, manage change, and repeatedly verify exploitable elements in the real environment.