The Strategic ShiftThe cyber kill chain is compressing
Frontier models are beginning to combine capabilities that once required different specialists, tools and handoffs.
The most important change in cybersecurity is not that AI can write code. It is that advanced models can increasingly move from discovery to reasoning, testing and adaptation in one continuous loop. The distance between finding a weakness and understanding how it could be exploited is shrinking.
For years, defenders benefited from friction. Subtle vulnerabilities demanded rare expertise. Reliable exploits required time. Scaling an operation required infrastructure and operators. Frontier AI lowers each of those barriers. It can inspect large codebases, maintain context across complex systems and continue working through failed attempts without fatigue.
"The defensive question is no longer whether AI will find more vulnerabilities. It is whether we can verify, prioritize and remediate them before the same capability is weaponized."
That compression changes attacker economics. Vulnerability discovery becomes cheaper. Lower-severity flaws can be chained into critical paths. Social engineering becomes more personalized. Agentic systems can run multi-step campaigns with less operator attention. None of these techniques is new; what is new is the speed and coordination with which they can be assembled.
The same capabilities can strengthen defense. Models can review code, propose patches, generate tests and help modernize legacy software. The outcome will depend on whether organizations connect powerful models to disciplined identity, evidence, approval and remediation processes.
Project GlasswingA controlled head start for defenders
Anthropic's program offers an early view of what happens when vulnerability discovery becomes abundant.
Project Glasswing began by giving vetted technology, infrastructure and security organizations controlled access to Claude Mythos Preview. Anthropic describes Mythos as a frontier model with unusually advanced cybersecurity capability and says it can outperform all but the most skilled humans at finding and exploiting software vulnerabilities.
In its first public update, Anthropic reported that roughly 50 partners had collectively found more than 10,000 high- or critical-severity vulnerabilities. Several partners said their rate of bug discovery increased by more than tenfold. The program later expanded toward roughly 200 organizations across more than 15 countries, including providers in power, water, healthcare, communications and hardware.
The figures are vendor-reported and should be evaluated accordingly. Even so, the operational lesson is difficult to ignore. Anthropic estimated that 530 high- or critical-severity bugs had been disclosed, another 827 were confirmed and awaiting disclosure, and only 75 of the disclosed bugs had been patched at the time of its May update.
Discover
AI expands the candidate pool.
Verify
Humans prove impact and remove noise.
Disclose
Owners coordinate safely.
Patch
Engineering creates and tests fixes.
Deploy
Operations verify risk reduction.
This is the inversion security leaders must plan for: discovery is no longer the scarce resource. Verification, coordinated disclosure, engineering capacity and deployment discipline become the constraints. An AI scanner attached to an overloaded queue creates activity; a governed remediation engine reduces risk.
The ImplicationsFive capabilities change the risk equation
Frontier AI does not replace the cyber kill chain. It compresses it — and places more pressure on identity, permissions and execution controls.
Vulnerability discovery
Models can inspect enormous codebases, connect unusual behaviors and surface subtle flaws missed by conventional scanning or time-constrained manual review.
Exploit chaining
Reasoning models can connect lower-severity weaknesses into a high-impact path, closing the gaps that once required several specialists.
Autonomous campaigns
Agentic systems can coordinate reconnaissance, targeting, execution and adaptation across long-running workflows with less human attention.
Synthetic social engineering
Personalized content reduces the cost of phishing, impersonation, business-email compromise and deepfake-enabled fraud.
Enterprise agent abuse
An over-privileged agent that can read data, call APIs and execute tools becomes a powerful path to leakage, unsafe change and lateral movement.
The common denominator is autonomy. Defenders must secure more than model output: the identities models assume, the tools they can invoke, the context they consume and the environments in which their actions execute. High-consequence actions should remain evidence-backed, logged and subject to explicit human authorization.
This also changes how security teams measure success. Finding volume is not the objective. The useful metric is time from validated finding to verified risk reduction, segmented by asset criticality, business service and software ownership.
The CISO ResponseBuild a remediation engine, not a finding factory
Organizations do not need access to Mythos to prepare for Mythos-class capability. The immediate work is architectural and operational.
Now
Inventory the blast radius map where models and agents touch code, credentials, sensitive data and production tools. Define which actions require human approval.
Next
Prove one governed workflow. Select a critical application. Exercise discovery, verification, ownership, patch generation, testing, rollback and disclosure end to end.
Scale
Expand with evidence. Automate low-risk evidence collection and testing only where the pilot reduces time-to-safe-state without increasing unverified noise.
The goal is a permanent defender advantage
Attackers may gain speed, but defenders retain structural advantages: ownership of the code, access to internal telemetry, authority over identity and the ability to remove systemic weakness. Those advantages matter only when security and engineering can coordinate them at machine speed without sacrificing judgment.
Project Glasswing should be read as both warning and blueprint. The warning is that advanced cyber capability will diffuse. The blueprint is controlled access, rigorous evidence, shared learning and relentless focus on patching. Organizations that build those disciplines now will not simply detect faster; they will become harder to attack.
