Cybersecurity company Chainguard has warned that the rapid advancement of artificial intelligence (AI) tools capable of discovering software vulnerabilities could trigger an unprecedented wave of security disclosures, potentially overwhelming organisations that are already struggling to install existing security updates.

The warning comes as cybersecurity researchers increasingly use advanced AI models to identify weaknesses in widely deployed software, develop working exploits and generate patches at a speed that traditional security teams may find difficult to match.

In an update on its Athena security research initiative, Chainguard disclosed that its systems have processed more than 40,000 vulnerability findings since June 2026, with approximately 42 per cent classified as critical or high severity.

The company said the findings span more than 500 software projects and have already resulted in the production of over 2,000 security patches.

However, Chainguard explained that identifying vulnerabilities and generating fixes are no longer its biggest challenges. The more difficult task is responsibly disclosing the discoveries, coordinating with software maintainers and ensuring that organisations receive security updates before attackers exploit the weaknesses.

The company warned that the cybersecurity industry could soon face a significant imbalance between the speed at which AI discovers vulnerabilities and the ability of organisations to fix them.

Chainguard Begins Disclosing Its First 50 Vulnerability Findings

Chainguard announced plans to begin publishing its first batch of approximately 50 Athena vulnerability discoveries, marking the beginning of a broader disclosure programme.

The initial findings were deliberately selected to avoid immediately exposing the most dangerous security weaknesses.

According to the company, the first 50 cases involve what it describes as silent fixes. These are vulnerabilities that software developers have already corrected in newer versions of their projects, sometimes years ago, without formally documenting them through the Common Vulnerabilities and Exposures (CVE) system.

Consequently, organisations running older versions of the affected software may remain vulnerable without receiving alerts from conventional security scanners.

Chainguard said this situation emerged repeatedly during Athena’s investigations because AI models examine software versions that organisations actually deploy, rather than focusing exclusively on the latest releases.

The company identified two major problems associated with these vulnerabilities.

First, software maintainers are often reluctant to backport security fixes to older versions, particularly when the relevant code has been completely rewritten or the affected development branch is no longer maintained.

Second, conventional security advisory systems may not identify these weaknesses because the vulnerabilities were already corrected in newer releases without receiving formal CVE identifiers.

To address the problem, Chainguard plans to publish the corresponding security patches, provide detailed vulnerability advisories and distribute corrected versions through its Chainguard Libraries platform.

The patches will be made available as publicly accessible files under the original software projects’ licences.

The company stressed that it does not intend to take ownership of the affected projects, create competing software versions or replace existing maintainers. Instead, the patches are intended to help organisations that cannot immediately upgrade their software.

Its security advisories will also include affected version ranges and technical details identifying the specific modules, classes and functions involved.

Chainguard Seeks Authority to Assign CVE Numbers

As part of its wider security initiative, Chainguard disclosed that it is working towards becoming a CVE Numbering Authority (CNA) in partnership with the United States Cybersecurity and Infrastructure Security Agency (CISA).

The proposed arrangement would allow the company to assign CVE identifiers to qualifying open-source vulnerabilities discovered through Athena.

The scope would primarily cover vulnerabilities that have already been corrected by upstream developers without receiving formal identifiers, as well as weaknesses affecting abandoned projects without maintainers capable of assigning CVEs.

The company said the initiative supports Gold Eagle, a federal cybersecurity clearinghouse established following an executive order in June.

Gold Eagle is intended to help ensure that important vulnerability information reaches operators of critical infrastructure, including organisations responsible for essential services that may require additional time to implement security updates.

Chainguard acknowledged that assigning CVE numbers to previously undocumented vulnerabilities could initially increase the number of security alerts appearing in organisations’ scanning systems.

However, it argued that identifying previously hidden weaknesses is preferable to allowing vulnerable software to remain undetected.

The company also said it would apply strict verification requirements before assigning identifiers.

Every reported vulnerability must be reproduced and supported by technical evidence rather than relying solely on AI-generated claims. Researchers must establish the affected software versions, identify the relevant code changes and notify maintainers before proceeding.

Where a project has an active and reachable maintainer, the developer’s own CNA remains responsible for assigning the vulnerability identifier.

Chainguard described its proposed role as temporary support for gaps in the existing disclosure system rather than an attempt to take control of open-source security reporting.

AI Model Discovers Multiple Critical Vulnerabilities in Six Hours

Chainguard also disclosed findings from its participation in Anthropic’s Project Glasswing, an initiative involving organisations using advanced AI capabilities to strengthen software security.

The company said its participation has given its Athena research programme access to Anthropic’s Mythos model, which it is using to examine software projects and identify previously undiscovered vulnerabilities.

One particular experiment demonstrated the potential scale of AI-assisted vulnerability discovery.

According to Chainguard, a single examination of a widely used open-source project lasted approximately six hours and produced more than a dozen validated vulnerabilities.

Most of the discoveries reportedly involved pre-authentication remote code execution flaws, which could potentially allow attackers to execute malicious commands without first obtaining login credentials or requiring assistance from a system operator.

The company said every reported vulnerability was reproduced three times inside a clean virtual machine before being accepted as a confirmed finding.

The project was not an abandoned or poorly maintained software package. Rather, it was described as one of the most heavily tested and closely examined projects in the open-source ecosystem, having undergone years of expert security reviews and extensive fuzz testing.

Despite those previous security efforts, the AI-powered investigation reportedly identified multiple serious weaknesses within a single afternoon.

Chainguard said the discoveries would be processed through Akrites, a coordinated vulnerability disclosure organisation established by the Linux Foundation in June.

The company declined to identify the affected software project while the disclosure process remains ongoing.

Security Researchers Warn About AI Sandbox Vulnerabilities

Another major concern raised by Chainguard involves the security environments used to execute AI agents during vulnerability research.

The company explained that discovering and validating software vulnerabilities often requires AI systems to generate and execute potentially dangerous code repeatedly.

Without adequate isolation, such activities could expose the underlying infrastructure to compromise.

Chainguard referenced previous incidents involving AI agents escaping security restrictions during testing and accessing systems outside their authorised environments.

It also cited research by Pillar Security, which documented sandbox escape vulnerabilities involving several AI coding tools, including Cursor, Codex, Gemini CLI and Antigravity.

According to Chainguard, some of these incidents demonstrated that attackers or malicious agents do not necessarily need to break directly through a sandbox.

Instead, they can exploit activities that are permitted inside the restricted environment, such as creating files that are subsequently executed by trusted processes outside the sandbox.

The company said this highlights the difference between having a sandbox and properly securing the entire environment surrounding it.

To address the challenge, Chainguard operates Athena inside hardware-isolated microvirtual machines.

Each research task receives its own isolated environment, including a separate operating system kernel and a freshly constructed filesystem.

Network access is restricted through the hypervisor, with connections permitted only to approved software package registries.

The company also said that attempts to establish unauthorised external connections automatically terminate the virtual machine.

Additionally, sensitive credentials are never placed inside the research environment.

Chainguard said it also tests the security of its sandbox infrastructure by directing AI agents to search for weaknesses in the isolation mechanisms themselves.

How Athena Uses Multiple AI Agents to Discover Vulnerabilities

The company further explained that Athena’s research framework uses multiple AI agents working simultaneously to examine different sections of software projects.

Rather than relying on one AI model to maintain the entire investigation in its memory, the system divides research activities into smaller tasks.

Each agent receives a specific assignment and information about previously discovered findings, allowing it to continue investigations without depending on lengthy conversation histories.

Chainguard said an important principle of the system is that an AI model cannot independently declare a vulnerability valid.

Instead, the host environment must reproduce the reported security weakness against a clean software installation.

A finding is accepted only when the vulnerability can be demonstrated before a patch is applied and the corresponding security test succeeds after the correction.

The company believes this verification process significantly reduces false positives and prevents AI-generated assumptions from being treated as confirmed vulnerabilities.

It also highlighted the importance of recording unsuccessful investigations because understanding where an AI agent has already searched can help other agents avoid repeating the same work.

Chainguard argued that the quality of an AI model’s decisions about where to investigate is more important than simply increasing the number of agents operating simultaneously.

Chainguard Plans to Open Its Security Infrastructure

Beyond Athena, the company disclosed that it has spent the past year rebuilding its software development infrastructure around three major technologies: isolated microvirtual machines, an AI-powered vulnerability research framework and a redesigned continuous integration system.

The microVM technology provides a secure environment for executing potentially dangerous code.

The research framework enables multiple AI agents to investigate software vulnerabilities, while the redesigned continuous integration system allows code changes to be tested and verified within the development process.

Chainguard said its internal infrastructure now handles more than ten times the volume of CVE-related work it previously managed.

The company plans to make these technologies available to external organisations over the coming quarters.

The tools are intended to support businesses running untrusted AI workloads, software developers seeking to identify vulnerabilities and organisations struggling with inefficient security testing processes.

Chainguard Warns Existing Security Systems May Be Overwhelmed

In its concluding assessment, Chainguard warned that the biggest cybersecurity challenge may not necessarily be the discovery of new vulnerabilities but the industry’s ability to respond to them.

The company noted that many organisations are already months behind in installing patches for vulnerabilities that have been publicly documented for years.

Introducing thousands of additional findings into an already overloaded security ecosystem could make the situation significantly worse.

It argued that faster vulnerability discovery and disclosure will only protect organisations capable of implementing security updates quickly.

However, many businesses, particularly operators of critical infrastructure, cannot respond at the same speed as advanced AI systems can discover security weaknesses.

Chainguard therefore believes that long-term cybersecurity protection must go beyond continuously identifying and patching vulnerabilities.

The company is advocating a transition towards software architectures that are secure by design, where common vulnerability categories such as memory corruption, insecure deserialisation, injection attacks and confused deputy problems are prevented from occurring in the first place.

It argued that repeatedly patching weaknesses in software built on vulnerable architectural foundations will not provide a permanent solution.

Instead, the future of software security should involve building new foundations that make entire classes of vulnerabilities impossible or significantly harder to introduce.

Chainguard said its experience processing thousands of vulnerability findings gives it insight into the weaknesses that matter most, while its AI-powered testing infrastructure could help verify whether redesigned software architectures eliminate those problems.

The company concluded that although slowing the development of increasingly powerful AI systems could provide additional time for security teams, strengthening the software infrastructure on which businesses and critical services depend is equally important.

Its central warning is that the cybersecurity industry is approaching a massive increase in vulnerability discoveries, but the systems responsible for identifying, distributing and fixing those weaknesses are already struggling to keep pace.