How AI turned cybersecurity into a race against time

Cybersecurity is undergoing a structural transformation — one defined not by a new malware family or a single disruptive technology, but by speed.

The pace at which vulnerabilities are discovered, weaponized, and exploited is converging with the speed at which AI systems can reason about code, infrastructure, and entire attack surfaces. Anthropic’s preview of Claude Mythos, and the decision not to release it publicly, marks a visible inflection point in that transformation.

For close observers, this shift has been visible for some time. But over the past two months it stopped being a forecast and became a measurable reality. Anthropic's decision not to release its Claude Mythos Preview model, the first published results from its Project Glasswing initiative, and OpenAI's rapid arrival with GPT-5.4-Cyber and later GPT-5.5(-Cyber), and its Daybreak platform have together confirmed what many security leaders suspected: vulnerability discovery and exploitation are accelerating beyond the pace of human-led security operations.

Importantly, this trend is no longer limited to a handful of frontier AI labs. Major industry players are increasingly positioning themselves around AI-driven security operations and accelerated cyber defense. Microsoft has highlighted this direction through its MDASH initiative, while CrowdStrike has introduced Quiltworks as part of its vision for AI-enabled security workflows.

Together, these developments point to a broader industry consensus: speed, automation and AI-assisted decision-making are becoming central to modern cyber defense.

The accelerating arms race

In April 2026, Anthropic previewed Claude Mythos, a general-purpose frontier model with cyber capabilities so advanced that the company chose not to release it publicly and publish a safer, more guardrailed Fable 5 model about 2 months later. This was not an act of simple caution. It was a signal in an accelerating arms race.

According to Anthropic, its AI systems have already identified high-severity vulnerabilities across virtually every major operating system and widely used web browser in use today. Beyond sheer coverage, the more unsettling aspect is the depth of what the system could do once those vulnerabilities were found. The reported capabilities extend far beyond basic scanning or static analysis.

Claude Mythos and related systems were able to autonomously complete the following:

Chain vulnerabilities into full attack paths
Develop complex exploit techniques
Escape sandboxed environments
Produce working privilege escalation exploits
Complete exploit development in less than 24 hours in some cases

In other words, the model did not just find weaknesses, it understood how to weaponize them with a level of creativity and speed that begins to rival top-tier human offensive teams.

What makes this even more consequential is how these capabilities emerged.

Evolution. Automation. Agentic autonomy.

The model was not explicitly trained to conduct offensive cyber operations. According to Anthropic’s red-teaming teams, the behaviors surfaced as a byproduct of generalized improvements in code reasoning, logical analysis, and agentic autonomy. The same ingredients used to improve software engineering productivity, automate debugging, or assist with code migration are now amplifying offensive cyber effectiveness. This dual-use reality is not a hypothetical scenario; it is observable behavior in frontier systems. Vulnerability discovery and exploitation are accelerating beyond the pace of human-led security operations.

There are profound implications on both the defensive and offensive sides.

On one hand, such models can dramatically strengthen defenders. They can continuously scan complex environments, identify misconfigurations and zero-days at scale, generate patches or compensating controls, and validate security architectures at machine speed. On the other hand, the same capability in the hands of a determined attacker collapses the cost and expertise required to find and exploit critical flaws.

Crucially, this is no longer a single-vendor story. Within weeks, OpenAI released GPT-5.5 and then GPT-5.5-Cyber, a variant tuned to be more permissive on security tasks for vetted defenders, followed by Daybreak, a platform that embeds frontier models directly into the software development lifecycle.

The line between “security tool” and “offensive capability” becomes disturbingly thin.

First, the vulnerability discovery-to-exploit timeline is getting fundamentally compressed. Historically, the gap between a vulnerability being discovered and exploited in the wild was measured in months. Over time, it shrank to weeks, then days. With AI-augmented vulnerability research, we are moving into a world where that window can be measured in hours. Security programs built around periodic scanning, scheduled patch cycles, and manual ticket queues are increasingly misaligned with this reality. Even mature organizations may be continuously exposed while still technically compliant with their own processes.

Second, advanced exploitation is being democratized. Capabilities that once required elite offensive researchers or nation-state investment are becoming more accessible. AI lowers the barrier to developing complex exploits, chaining seemingly minor bugs into full compromise, and weaponizing zero-days. That does not mean every script-kiddie suddenly becomes a nation-state actor, but it does mean the overall baseline of adversary capability is rising, and it is rising fast.

Third, AI versus AI is becoming the default operating condition. Human-only security operations centers cannot operate at machine speed. Detection, triage, and response processes that depend on manual investigation and hand-crafted correlation rules will fall behind. AI-augmented defense is no longer a differentiator; it is becoming a minimum requirement. The relevant question is no longer “Have we identified the vulnerability?” but “Can we decide and act fast enough before it is exploited?”

The Glasswing results moved the conversation from a vendor's self-assessment to externally observable evidence. Several partners reported their bug-finding rate increasing more than tenfold. Cloudflare found roughly 2,000 bugs (400 of them at high- or critical-severity) with a false-positive rate its team considered better than human testers. Mozilla found and fixed 271 vulnerabilities in a single Firefox release, more than ten times what it surfaced in the prior version using an earlier model. Palo Alto Networks shipped roughly five times its usual number of patches; Oracle reported finding and fixing flaws several times faster than before. Since June 2, 2026, the initiative has grown by approximately 150 participating companies, further expanding the scale of vulnerability discovery and remediation efforts.

Beyond its partners, Anthropic scanned more than a thousand open-source projects, surfacing an estimated 6,202 high- or critical-severity vulnerabilities, with a 90.6% true-positive rate across an independently triaged subset.

Unauthorized access, rising risks and governance spiraling out of control

The reported incident involving unauthorized access to a restricted Claude Mythos environment (despite not confimed as of today) illustrates a broader concern. The claims about the hack could not be verified by Anthropic, but the implication stands: if such access pathways exist, even temporarily or indirectly through a third-party vendor environment, they represent a meaningful security exposure.

According to the reports, a small group of unauthorized users may have accessed an unreleased model through an external vendor integration. Even assuming the situation was contained, the scenario itself remains significant from a risk perspective. A system described as capable of autonomously identifying vulnerabilities and generating sophisticated exploits, if exposed in an uncontrolled context, would constitute a high-leverage asset for adversaries.

In such a scenario, threat actors could potentially attempt to extract sensitive training data, manipulate outputs, or leverage the model as an acceleration layer for cyber operations such as phishing campaigns, malware development, or automated reconnaissance. The core concern is that advanced AI systems can materially lower the barrier to executing large-scale, high-impact attacks against enterprises, governments, and critical infrastructure.

Importantly, this is not a risk confined to a single vendor or model. It reflects a systemic challenge across the rapidly evolving ecosystem of high-capability AI systems, many of which are being integrated into production environments at speed. As a result, governance, access control, and third-party security assurance become as critical as model performance itself.

This is why governance, access control, and vendor security suddenly matter as much as model performance. Strong authentication and authorization, isolated execution environments, continuous monitoring, and rigorous vendor assessments are not optional. Organizations experimenting with or deploying high‑risk AI models, especially those handling sensitive data or making high‑impact decisions, need explicit policies that define who can access them, under what conditions, and with what guardrails. The same rigor historically applied to sensitive cryptographic keys or offensive security tools now needs to be applied to powerful AI systems.

A practical crisis playbook for the C-suite

Atos’s portfolio with Exposure Management Readiness Assessments, continuous Exposure Management program design and build, Security Program (AI) Transformation, Zero Trust Architecture, Zero‑day Crisis Readiness, and Security for AI delivers the practical controls, testing, and program changes needed to compress the time from detection to remediation. These offerings leverage embedded AI‑specific controls, continuous testing, and governance into your security program so that risks are identified earlier, prioritized intelligently, and addressed with a fast, measurable response. By combining readiness assessments, AI security transformation, Zero Trust, and crisis playbooks, we create automated, scalable detection‑to‑remediation paths that match pace with increasingly capable models.

But the same AI capabilities that threaten to outpace human defenders can also be harnessed to restore balance.

Organizations that adapt successfully will share several characteristics: continuous exposure visibility across infrastructure and applications, AI-augmented security operations that can detect and respond at machine speed, deliberate reduction of attack surface through secure-by-design architectures, automated validation of controls, and explicit strategies to secure AI systems themselves. These organizations will not eliminate risk, but they will reduce exposure structurally and regain control of time to decision, time to remediation, and time to containment.

The acceleration of AI-driven vulnerability discovery has been visible to close observers for some time. What developments like Claude Mythos, Project Glasswing, and the associated access incident confirm is that the timeline has moved forward dramatically. For security leaders, this is less about reacting to a single breakthrough and more about acknowledging a new operating reality: the future of cybersecurity will not be defined by how many tools an organization deploys, but by how quickly it redesigns its security program to operate at AI speed.

The arms race is already underway. A key question to ask ourselves is whether security programs can evolve fast enough to keep up.

>> Connect with us to learn more about how Atos can help your organization reconstruct a future-fit security program today.

>> Learn more about Atos is helping global players build sustainable and scalable security solutions for their business: https://atos.net/en/services/cybersecurity

>> Check our new cybersecurity whitepaper on Adaptive Cyber Resilience in the Age of AI

Posted: 10/06/26