# The AI Vulnerability Storm, Detailed

A briefing on where state of the art AI vulnerability discovery stands today, how it is evolving and what Sonatype recommends you do about it. **~60 AI-discovered CVEs** are public to date; **thousands** more are announced and awaiting vendor patches.

**AI-discovered CVEs · Public**

60–65

**Vendor-credited AI attribution in advisory or CVE record**

**AI-discovered · Under Embargo**

3,200+

**Vendor-claimed, disclosure pending through Q3**

**Ratio · Embargoed ÷ Public**

~53×

### The structural gap now defining the landscape

**Active AI Programs**

8: AISLE, Anthropic, XBOW, Horizon3, OpenAI, RunSybil, ZeroPath, AIxCC

### What this means for you

Your upstream suppliers — OSS projects, browsers, kernels, commercial vendors — are now being audited, at scale, by machines that do not sleep. Over the next two quarters, the rate at which fixes land in the components you depend on is going to change step-function.

### State of the art · today

Four production-grade programs — **AISLE**, **Anthropic** (Claude Opus 4.6 → Mythos), **XBOW**, and **OpenAI Codex Security** [21](/content/resources/research/ai-vulnerability-storm#ref-21/index.html) — are already shipping CVEs against OpenSSL, curl, Firefox, FreeBSD, the Linux kernel, Apache ActiveMQ and Microsoft cloud services. Time-to-exploit has collapsed to under a day. The volume of vulnerabilities being _discovered_ is on track to overwhelm the industry's capacity to **apply patches** — remediation, not disclosure, is the real bottleneck.

### How it is evolving

Anthropic's April **Mythos Preview** [13](/content/resources/research/ai-vulnerability-storm#ref-13/index.html) claims thousands of findings with >99% still embargoed under SHA3-224 hash commitments, coordinated through **Project Glasswing** with roughly 40 early-access vendors. The first mandatory disclosures open **July 7** and close around **September 5, 2026**. Mythos-class capability will proliferate to other frontier models within months, and open-weight models within a year [13](/content/resources/research/ai-vulnerability-storm#ref-13/index.html) — this is the first wave, not the last.

### What Sonatype recommends

Treat Q3 as a **remediation surge**, not business-as-usual. Control the ingress — every package entering the enterprise should pass a policy gate, no CVE required. Put real-time intelligence in front of every coding agent. Gate every build with SCA, SBOM, and release policy. The **90-day plan** at the end of this report maps these moves to _Firewall_, _Guide_, and _Lifecycle_.

### A note on terminology

#### What is a hash commitment?

When Anthropic says Mythos found "thousands of zero-days" but refuses to disclose them, how does anyone verify the claim without the vulnerabilities leaking? The cryptographic answer is a **hash commitment** — also called a commitment scheme.

The researcher takes the full vulnerability report plus proof-of-concept exploit, runs it through a one-way hash function (Anthropic uses `SHA3-224`), and publishes _only the resulting hash_ — a 56-character fingerprint that reveals nothing about the bug. Later, when the vendor has patched, the researcher publishes the original document. Anyone can re-run the hash and confirm it matches — proving the researcher knew about the vulnerability on the earlier date without tipping off attackers.

In practical terms, it's a **sealed envelope**: the hash is the wax seal, the PoC inside stays private until disclosure, and the cryptography guarantees the envelope can't be swapped.

### Worked example

| Step | Description |
| --- | --- |
| 01 | Vulnerability report + PoC (kept private) |
| 02 | Public hash commitment (published Apr 7) |
| 03 | Vendor patches, researcher reveals |

>99% of Mythos findings still under commitment, by Anthropic's own count.

135 days Max window from vendor report to mandatory disclosure (90 + 45).

### The ground was already shifting

CVE volume tripled before AI-discovery scaled

#### The baseline was already breaking

Before any AI program shipped a CVE, the public catalog was already buckling under sheer volume. CVE submissions to the National Vulnerability Database **increased 263% between 2020 and 2025** [26](/content/resources/research/ai-vulnerability-storm#ref-26/index.html) and Q1 2026 is running roughly **one-third higher** than Q1 2025. The AI-discovery surge documented in this report is additive — it is piling onto a system that was already losing ground.

#### The institution that scored them is retreating

On **April 15, 2026**, NIST announced that the NVD will no longer enrich every submitted CVE [26](/content/resources/research/ai-vulnerability-storm#ref-26/index.html). Going forward, only CVEs in CISA's KEV catalog, software used by the U.S. federal government, and EO 14028 critical software get analysis — everything else is marked _Not Scheduled_. NVD enriched ~42,000 CVEs in 2025 (a 45% productivity gain) and still couldn't keep up.

### Evidence ledger

Every claim in this report, traceable to a primary source

Fairly structured data appears regarding CVEs, including dates, discoverers, and findings across different security incidents. This raw data reflects references made throughout the **AI Vulnerability Storm** report, with witnessed and attributed incidents supporting the report’s findings.

#### Conclusion

The report suggests an overwhelming rise in AI-discovered CVEs that will radically reshape vulnerability management workflows and recommends adopting systematic approaches that prioritize remediation, real-time intelligence, and regulatory compliance in the evolving landscape of cybersecurity.