As the artificial intelligence landscape accelerates at an unprecedented pace, industry leaders and security experts are increasingly grappling with the profound gap between advanced technological capabilities and the defensive readiness of global infrastructure. Following the publication of Dario Amodei’s essay, We Must Pace the Frontier, cybersecurity circles have turned their attention toward a stark realization: while managing the raw power of frontier AI models in laboratories remains a critical challenge, the immediate danger lies in how these systems interact with decades of legacy software underpinning modern critical infrastructure.
The security community has spent the summer examining what these advanced models can uncover. The resulting consensus points to a dual-layered pacing problem. On the frontier end, researchers and governments face the daunting task of regulating the development of powerful foundational models. However, the consumption end—comprising disclosure processes, patch pipelines, and the underlying software itself—presents an even more urgent operational challenge. Software powering the power grid, water systems, and hospital networks largely runs on decades-old architecture with familiar, exploitable design flaws. While frontier pacing buys time at the development stage, the real battleground rests in how effectively the software ecosystem can absorb what these models ultimately produce.
The First 50 and the Reality of Silent Fixes
Since June, the Athena initiative has been actively ingesting security findings at an extraordinary scale. Intake numbers reflect more than 40,000 total findings across more than 500 projects, with 42 percent categorized as critical or high severity, culminating in the production of over 2,000 patches. Yet, the primary bottleneck has never been the generation of fixes; rather, it is the unprecedented scale required for responsible vulnerability disclosure.
To safely test the ingestion and distribution pipeline, the initiative rolled out its first batch of approximately fifty patches. Notably, these initial releases avoided high-severity, live pre-authentication remote code execution flaws, which instead follow strict embargo protocols through shared disclosure bodies. Instead, the first wave focused entirely on silent fixes—vulnerabilities that were patched at the code’s head, sometimes years prior, but never formally accompanied by a Common Vulnerabilities and Exposures (CVE) identifier.
This pattern emerged rapidly within Athena, revealing that a significant portion of model-discovered flaws affect older software versions rather than current releases. Because AI models scan what is actively deployed in the wild, and because deployed systems frequently run legacy code, they expose gaps that standard vulnerability scanners overlook. Standard upstream processes typically do not merge patches into outdated code branches, leaving advisory layers without data since CVEs are rarely filed for issues already resolved at head.
To bridge this gap for active projects lacking pathways for legacy patches, the initiative adopted a structured approach. Plain patch files are published in a dedicated repository under original project licenses, serving as an interim solution for operators who cannot immediately upgrade. Simultaneously, OSV vulnerability feeds with precise version ranges and symbol-level details are published so that partner scanners can surface the data. Furthermore, patched versions are shipped directly into managed libraries for enterprise customers as soon as advisories go live, minimizing friction while maintaining transparency.
Becoming a CNA to Meet the Scale of Discovery
Historically, a silent fix carried minimal risk because discovering the underlying vulnerability from a commit log required specialized human expertise and considerable time. The advent of frontier AI models fundamentally alters this dynamic. An advanced model can extrapolate a working exploit against unpatched versions from a public fix commit within a single afternoon. Consequently, the vulnerability exists regardless of whether it is publicly documented, leaving the security community with a single imperative: ensuring defenders receive the information as well.
To address the lack of tracking mechanisms for silent fixes, Chainguard has initiated steps to become a CVE Numbering Authority (CNA) in partnership with the Cybersecurity and Infrastructure Security Agency (CISA). This CNA scope focuses specifically on open-source vulnerabilities processed through Athena where upstream projects fixed flaws without identifiers, or where maintainers are no longer available to assign them.
While this expansion will inevitably introduce additional noise into enterprise vulnerability scanners as long-overlooked legacy versions light up, security advocates argue that visibility is preferable to a false sense of security born of silence. To mitigate false positives, advisories are tightly scoped with precise introduction and fixed ranges across release branches, providing reachability scanners with the module, class, and function-level details necessary to filter irrelevant alerts. Crucially, this role remains transitional rather than territorial; if a maintainer is reachable and a vulnerability exists in the latest version, the project’s own CNA retains assignment authority.
Managing Disclosures and Infrastructure Risk
For vulnerabilities affecting current versions of maintained projects, workflows are routed through established collaborative bodies. Findings involving reachable maintainers are processed through Akrites, the shared disclosure organization launched by the Linux Foundation, while unmaintained projects fall under maintainer-of-last-resort frameworks coordinated within the same body. Meanwhile, high-priority findings impacting critical infrastructure are channeled through Gold Eagle, the federal clearinghouse established via executive order to ensure critical operators receive timely threat intelligence.
Operating a shared incident response team across numerous organizations introduces initial administrative friction compared to isolated responses, but stakeholders view it as an essential trade-off to protect maintainers from being overwhelmed by independent reports.
The Speed of Frontier Discovery
The urgency behind these structural shifts is underscored by real-world testing. Through participation in Anthropic’s Project Glasswing, security teams integrated advanced models, such as Mythos, directly into their own software development and shipping pipelines.
In a representative test involving a widely used open-source project—code that has undergone years of rigorous expert review, fuzzing, and heavy scanning—an AI model identified over a dozen validated vulnerabilities within a six-hour run. Most of these were pre-authentication remote code execution flaws requiring no credentials or operator action. To maintain rigorous accuracy, each finding was subjected to a strict verification rule: the AI model was not permitted to self-validate its discovery; instead, a clean virtual machine host had to independently reproduce the exploit multiple times before the finding entered the record.
This rapid discovery rate highlights a widening divergence between two opposing timelines. On one side, the software industry attempts to absorb a massive influx of vulnerabilities using disclosure processes that were already backlogged. On the other, the capabilities of automated exploit generation continue to advance toward a threshold where large-scale compromise of critical infrastructure becomes economically and technically trivial.
Sandboxes, Harnesses, and Operational Infrastructure
The operational mechanics required to safely run autonomous vulnerability discovery agents demand rigorous isolation boundaries. Generating and detonating working exploits against real code via autonomous agents presents inherent risks, as demonstrated by recent high-profile sandbox breakouts across various commercial AI developer tools. Often, these security failures do not stem from direct structural breaches of the containment container, but rather from systemic gaps—such as agents writing permitted files that are subsequently executed by trusted processes outside the boundary.
To prevent such incidents, discovery pipelines utilize hardware-virtualized microkernel microVMs. In this architecture, every task executes inside an isolated guest environment with a custom-built root filesystem, strict hypervisor-enforced egress limited exclusively to approved package registries, and a policy where any unauthorized connection instantly terminates the virtual machine without credentials ever entering the guest. Agents are deliberately pointed at these sandboxing layers to test their resilience, ensuring that every identified vulnerability first translates into a failing test case before a patch is developed and verified.
The automated harnesses driving these microVMs draw conceptual inspiration from advanced human-led vulnerability research, where complex exploit chains are uncovered through iterative cycles of trial, failure, and redirection. By sharding these processes across multiple parallel agents, systems can maintain continuity through durable findings rather than relying on a single human working memory context window.
The Ultimate Endgame: Secure-by-Construction Foundations
Ultimately, faster disclosure pipelines and rapid patching schedules address only the symptoms of a deeper systemic vulnerability. The vulnerability classes consistently exploited by frontier models—such as memory corruption, deserialization flaws, injection vulnerabilities, and confused deputy problems—are structural consequences of how enterprise software stacks were originally designed. Organizations cannot simply patch their way out of fundamental architectural weaknesses.
For decades, the high cost of migration protected legacy platforms from widespread rewriting, as enterprises could rarely afford to replace foundational codebases like COBOL. However, the emergence of capable AI agents drastically lowers the economic barriers to large-scale code transformation.
The ultimate resolution to the escalating threat landscape requires transitioning toward new foundational architectures that are secure by construction, rendering dominant vulnerability classes impossible to write in the first place. While pacing the frontier through coordinated regulation and robust disclosure buys crucial time, true long-term security depends on removing the underlying targets entirely, breaking the cycle of perpetual remediation and establishing a resilient baseline for the future of enterprise software.
