Anthropic’s Mythos and Glasswing: Where Do Defenders Gain and Struggle?
Published
In our first Ask the Experts panel, we examine what happens when patching cannot keep up, and exposure becomes a built-in condition.
With Anthropic’s Mythos and Project Glasswing drawing attention to AI capabilities in vulnerability discovery, we seek to understand how this affects real-world security operations.
While Anthropic has introduced Project Glasswing to a select group of major industry players, we are curious to know what this means for the broader security ecosystem. How would the equation change when this capability reaches other players? What happens when threat actors access it? Is the industry prepared for what comes next?
Taken together, these concerns are not theoretical. Even with current-generation tools, teams are dealing with increasing backlogs. AI is finding vulnerabilities faster than teams can fix, while existing patching processes were not even built for this speed or volume.
In complex environments, this means a landscape where discovery keeps accelerating, but response fragments. Here’s what cybersecurity experts have to say.
Question: If AI can find vulnerabilities faster than organizations can fix them, what breaks first: patching models, security teams, or business operations?
Ram Varadarajan, CEO at Acalvio
It's the new forever war -- the race between autonomous vulnerability discovery and human remediation.
- Business operations are going to fail first, perhaps catastrophically, at the remediation bottleneck: AI can identify zero-days in milliseconds.
- This collapses the traditional, human-scale, weeks-long patch cycle, and forces untenable trade-offs between uptime and exposure.
- The remediation lag, driven by maintenance constraints and stability testing, is going to combine with trust erosion, with automated exploitation outpacing manual-paced defense, and pushing breach costs beyond what will be acceptable margins.
This failure will be strategic as much as operational: symmetrical defenses that try to match attacker speed are no longer viable. We have to pivot to bot-on-bot defense. Specifically, deception-centric models such as hypergame environments, and model-aware deception that misdirect attackers and create a verification gap, buying time for stabilization.
Morey Haber, Chief Security Advisor at BeyondTrust
If AI can find vulnerabilities faster than organizations can fix them, the first thing that breaks is not patching. It is not even the business. It is the security operating model itself.
We are entering a phase where vulnerability discovery, exploit generation, and attack orchestration occur at machine speed, collapsing the window between discovery and weaponization to hours or less.
Traditional security teams were never designed for this cadence. They operate on human triage, ticket queues, and risk prioritization models that assume time exists between exposure, impact, and remediation. That model is now in question.
Consider these fracture points based on the document, “The “AI Vulnerability Storm”: Building a ‘Mythos-ready’ Security Program” released by the Cloud Security Alliance.
- Security teams break first. Not because of incompetence, but because of labor and time to complete. The volume and velocity of AI-discovered vulnerabilities create a workload that exceeds human scaling limits.
- The findings in the report explicitly warns of burnout, overwhelmed teams, and the need for reserve capacity and lack of automation to complete assigned tasks.
- When the defenders cannot keep pace, patching becomes a liability.
- Patching models will break next. Not because patching is ineffective, but because it becomes overwhelming. When exploitation timelines compress to minutes or hours, the concept of remediation before compromise becomes aspirational.
- The report highlights that we can no longer assume a patch will be ready in time for remediation. This shifts the paradigm from prevention to containment and resilience.
- Finally, business operations break last, but with the greatest consequences. Once both the security function and patching lifecycle are saturated, risk accumulates and executive teams will become aware of the short comings.
- At that point, disruption is no longer a possibility it will have board level responsibilities warranting action.
Jason Schmitt, Chief Executive Officer at Black Duck
If this imbalance continues, what breaks first isn’t the AI models—it’s the system around them. In the real world, security teams are already overwhelmed by volume, not unaware of risk.
AI has dramatically lowered the cost and time required to discover vulnerabilities, while remediation still depends on people, change control, regression testing, and business prioritization. That gap is widening. When discovery outpaces execution, teams default to two unhealthy patterns:
- patch fatigue or
- selective blindness.
- Neither reduces risk.
What’s changing now is speed and asymmetry.
- Attackers don’t need perfect fixes or governance; they need one unpatched weakness.
- Defenders, on the other hand,
- have to assess impact,
- understand reach across dependencies, and
- coordinate fixes across teams that don’t share incentives or timelines.
Current patching models weren’t built for AI‑generated code churn or continuously evolving supply chains, and most organizations don’t have the context to know which vulnerabilities actually matter.
If nothing changes, business operations will feel the impact next:
- delayed releases,
- emergency freezes,
- rising technical debt, and
- security teams forced into gatekeeping roles that slow innovation rather than enabling it.
That’s where trust breaks down. What organizations should do differently now is shift focus from raw detection to execution leverage.
- That means prioritizing vulnerabilities based on exploitability and business exposure, not scores alone,
- embedding security earlier in development where fixes are cheaper, and
- putting governance around AI‑generated code so risk doesn’t accumulate invisibly.
The goal isn’t to match AI’s speed, it’s to apply judgment, context, and control where machines alone fall short.
Nicole Carignan, Senior Vice President, Security & AI Strategy, and Field CISO at Darktrace
AI has been accelerating vulnerability discovery faster than most organizations can validate, prioritize, patch, deploy, and verify fixes.
So, the first thing that breaks is not necessarily the security team or the business itself, but the operating model that assumes remediation can keep pace with identification. In many organizations, budget breaks alongside it, because security investment is still not keeping pace with AI adoption or the scale of monitoring and tooling, now required.
The real-world problem is that patching has always been constrained by people, process, uptime constraint, and operational risk. Many organizations cannot rapidly update legacy systems, industrial environments, or business-critical platforms without disruption.
That means faster discovery does not automatically make organizations safer. In many cases, it creates larger backlogs, more triage pressure, and a greater chance that exploitable issues remain open longer.
What changes next is the security posture organizations need to adopt. If AI continues to compress the gap between discovery and exploitation, companies cannot rely on CVE tracking and patching alone.
They need to be able to detect exploitation of vulnerabilities they do not yet know exist. That means more emphasis on
- monitoring,
- behavioral analysis,
- anomaly detection,
- autonomous investigation, and
- fast containment, especially in environments where patching is slow or infeasible.
Organizations also need to recognize that they are defending against more than software flaws alone. They have to defend against identity and credential theft, human error, insider threats, misconfigurations, misuse of AI tools, and AI systems that unintentionally or intentionally introduce new risks.
Organizations should stop treating vulnerability management as a closed loop ending in a patch. They need to
- prioritize accurate, advanced anomaly-based threat detection and autonomous containment just as aggressively as remediation, and
- invest in security resilience at the same pace they are investing in AI adoption. Otherwise, the discovery-remediation gap will widen faster than most teams can absorb.
Diana Kelley, Chief Information Security Officer at Noma Security
What breaks first isn't patching or even security teams in isolation. It's the operating model that connects discovery to remediation.
Mythos shows AI can now find and exploit vulnerabilities at a level that surpasses all but the most skilled human researchers, at scale. Mythos Preview has already identified thousands of previously unknown, high-severity vulnerabilities across major operating systems, browsers, and open-source infrastructure, many long-standing and undetected.
- One example: a 17-year-old remote code execution vulnerability in FreeBSD, CVE-2026-4747, exploited autonomously. That changes the timeline defenders are operating on.
Glasswing is currently restricted to roughly 40 major partners. That restriction buys us time to address remediation practices, which are already under strain, before autonomous discovery breaks them.
Right now, remediation is still gated by ownership ambiguity, testing cycles, uptime requirements, and limited engineering bandwidth. Security teams accumulate findings faster than they can translate them into safe, prioritized action. The backlog becomes unmanageable, and prioritization collapses under volume.
The bottleneck is the operating model. Fixing at AI speed means accepting more instability and more risk to revenue systems. Most organizations haven't been willing to make that trade-off, so we slow down remediation instead.
Current approaches optimize for discovery, not execution. We've built pipelines that generate more signal than our systems can absorb. What needs to change is structural. We have to reduce dependency on patching as the primary control:
- shrink attack surface,
- improve asset intelligence,
- pre-position compensating controls.
In practice, that means clear ownership for every internet-facing asset and defaulting to segmentation or rate limiting when patch SLAs can't be met.
We also have to stress test our own systems proactively, using the same capabilities our adversaries will, including chained exposures and agent deployments, which are becoming targets.
The organizations that adapt fastest will treat remediation as a systems design problem, not a ticketing problem.
John Gallagher, Vice President of Viakoo Labs at Viakoo
With organizations now managing 5 - 10 times more network-connected OT, IoT, and CPS devices than traditional IT systems, the first thing to break under accelerated AI-driven vulnerability discovery will be business operations that are reliant on these non-IT environments.
OT patching models are already fractured and inadequate. They remain largely manual or device-specific — consider FDA-regulated medical devices or manufacturing systems that need scheduled downtime for updates.
Unlike IT, which benefits from mature, automated patch management, the OT/IoT landscape — with more than 150,000 distinct operating systems — lacks scalable automated solutions, leave alone the autonomous capabilities that are needed to counter the rapidly emerging exploitations like those uncovered by Mythos.
Current security strategies focus largely on vulnerability discovery and risk prioritization — the "find and notify" approach — but they fall short on the operational realities of timely remediation.
However, without an autonomous patch deployment process, surfacing exploitable vulnerabilities will inevitably bring OT/IoT/CPS systems to a halt. This operational breakdown will compel a fundamental shift in security team structures, incorporating line-of-business managers who oversee OT systems and expanding governance to fully encompass these environments.
Mythos-driven exploits will stress credential and configuration management, demanding faster, more autonomous controls.
To meet this urgent threat, organizations must reframe OT patching as a continuous, autonomous process embedded within operational workflows — not a periodic project or an afterthought.
Immediate priorities include:
- investing in precise asset visibility,
- integrating automated OT remediation where feasible, and
- aligning security, IT, and OT teams around unified risk-reduction metrics.
- Without addressing these practical constraints, faster vulnerability detection will only magnify exposure and exacerbate risk, rather than mitigate it.
Joe Saunders, Founder and CEO of RunSafe Security
Patching models were already under strain, especially in critical infrastructure, where updates can take months or even years.
What AI-driven discovery changes is the scale. We’re seeing a surge of zero-day vulnerabilities that no security team can realistically keep up with.
What breaks next are the security teams themselves. They’re being forced into constant triage as the volume of exploitable findings outpaces their ability to validate and remediate them, creating a growing backlog of known, unpatched risk.
For operational technology and embedded systems, the challenge is even more acute. These environments often require physical access, certification, or planned downtime to patch, making rapid response impossible. The assumption that you can fix vulnerabilities before they’re exploited is quickly becoming untenable.
This is an inflection point for the industry. Security can’t rely solely on patching and has to focus on reducing exploitability even when vulnerabilities remain. That means adopting protections that make software harder to attack in the first place, so organizations aren’t forced to choose between operational risk and security risk.
Related
