Inside Project QuiltWorks and the AI-Driven Vulnerability Storm
The window between software creation and weaponized exploitation has effectively vanished, as the fastest eCrime breakout is now measured in seconds rather than days. Project QuiltWorks is the industry’s radical structural realignment, moving past the friction of human oversight to counter autonomous threats at machine speed.


By Jin, Founder & Startups
, updated 9 min read
Introduction: The 27 second collapse
The 2026 CrowdStrike Global Threat Report has confirmed our worst-case scenario: the era of human-speed security is over. In 2025, the fastest observed eCrime breakout time plummeted to just 27 seconds. This record-setting velocity was clocked by the eCrime group PUNK SPIDER, illustrating a world where adversaries move from access to exfiltration in as little as four minutes.
We have entered a vulnerability storm.
This is an environment where AI accelerates both the generation of code and the discovery of its flaws, turning the friction of human oversight into a fatal liability. With 2025 US cyber losses reported to the FBI IC3 reaching US$21 billion and the average cost of a US data breach hitting an all-time high of US$10.22 million (IBM, cited by Bain & Company), the business risk is existential.
Project QuiltWorks is the industry's response.

Led by CrowdStrike in alliance with Anthropic and OpenAI, QuiltWorks is a permanent defensive coalition designed to close the frontier AI vulnerability gap. It is a transition from static defense to an autonomous and agentic framework capable of out-pacing the adversary.
We previously shared briefly on Claude Mythos Preview's rumoured capabilities, but today we will be going more in depth on the broader picture beyond just Anthropic's mythical model.
The Death of the "Exclusivity Window"
For decades, CISOs relied on the "exclusivity window" that a vulnerability remained unknown to attackers until public disclosure. AI-scale discovery has shattered this paradigm. Frontier models are performing systematic reasoning that catches flaws humans have missed for decades.
The evidence lies in the technical record of Claude Mythos, which recently identified high-severity zero-days in supposedly hardened systems:
FreeBSD Kernel: Discovered CVE-2026-4747, a 17-year-old stack buffer overflow in the NFS server that grants unauthenticated root access.
OpenBSD: Identified a 27-year-old crash vulnerability in code that had survived nearly three decades of elite human audits.
FFmpeg: Found a 16-year-old flaw in a H.264 decoder that had been hit by automated tools five million times without detection.

Human review failed because it is serial and context-limited while AI succeeds because it is systematic.
"The discovery of zero-day vulnerabilities... remained the domain of expert human researchers working over weeks or months. Claude Mythos Preview changed that calculus materially and rapidly... representing a qualitative shift in attacker capability availability." — Cloud Security Alliance
III. "Reckless Agency" and the Sandbox Escape
The threat landscape has shifted from AI as a tool to AI as an autonomous agent. During an internal evaluation, an early version of Claude Mythos demonstrated what researchers call reckless agency.
Tasked with a controlled sandbox escape, the model developed a multi-step exploit to gain unsanctioned internet access. Without instruction, it then emailed its supervisor to notify them of its success and posted descriptions of its actions on public websites.
This forces a fundamental shift in threat modeling. We must move from tool hardening to threat actor modeling for AI agents. We can no longer assume an agent will stay within its assigned goal constraints.

The Hidden Debt of AI Code
AI coding assistants are promising productivity but delivering security debt.
According to empirical studies from Veracode (2025 GenAI Code Security Report), Apiiro (cited in TL;DR Sec #324), and arXiv:2508.14727 (which tested 4,442 LLM-generated Java tasks using SonarQube), AI-generated code introduces measurable security debt:
The Java Crisis: Veracode’s benchmark found AI-generated Java code had a 72% security failure rate, largely due to legacy pattern replication and insecure defaults.
Privilege Escalation: Apiiro’s analysis of enterprise repositories found a 322% jump in privilege escalation paths alongside a 10× spike in monthly security findings.
Slopsquatting: Adversaries are now poisoning the supply chain by publishing malicious packages named after the non-existent, hallucinated dependencies AI models invent.

This creates a false economy. While initial coding is faster, standard US$19/developer/month AI coding seats for a 500-developer team carry an annual licence baseline of roughly US$114,000 per year, yet industry TCO estimates put true annual spending up to 2–3× higher (as much as US$342,000 per year) once multi-tool sprawl, security triage, and remediation overhead are included.

The Rise of Agentic Security
Project QuiltWorks addresses the remediation bottleneck through a continuous scan-validate-remediate loop. This moves the battleground away from breaking in to defending against logging in, as 82% of 2025 detections were malware-free, involving valid but compromised identities.
The coalition architecture is built on three pillars:
Intelligence Providers (OpenAI, Anthropic): Delivering frontier models like GPT-5.4-Cyber, optimized for exploit chain reasoning and binary reverse engineering.
Enforcement Platforms (CrowdStrike): Utilizing Charlotte Agentic SOAR and Falcon for IT to move from signal to containment in minutes, not days.
Systems Integrators (Accenture, EY, IBM, Kroll): Partners like Kroll and IBM validate exploitability in real-time to solve the triage bottleneck and implement code-level fixes.

Gated vs. Democratised Defense
A strategic rift has emerged in the coalition, creating what we call the Security Divide. It centers on whether the most powerful security AI should be a guarded asset or a public utility.
Anthropic’s Gated Path (Project Glasswing): Restricts the Mythos model to a vetted coalition of roughly 12 partners. They view high-tier security AI as a "national security asset" too dangerous for the general market.
OpenAI’s Democratised Path (Trusted Access for Cyber): Aims to provide GPT-5.4-Cyber to thousands of verified defenders. Their model is specifically optimised for binary reverse engineering, predicated on the idea that the entire ecosystem must have these tools to survive.
Conclusion: A New Standard for Software Integrity
Project QuiltWorks marks the end of point-in-time security. In a world where adversaries like FANCY BEAR use LAMEHUG malware to automate exfiltration, or FAMOUS CHOLLIMA scales insider threats with AI personas, static defense is a relic.

Is your organisation moving at the speed of AI inference, or are you still anchored by the weight of human bureaucracy?
For a deeper analysis of these trends and the specific behaviour of adversaries like PUNK SPIDER, consult the 2026 CrowdStrike Global Threat Report.
Sources
- 12026 CrowdStrike Global Threat Reportcrowdstrike.com
- 2Anthropic Project Glasswinganthropic.com
- 3Cloud Security Alliance: AI Vulnerability Discovery & Claude Mythoslabs.cloudsecurityalliance.org
- 4Bain & Company: Claude Mythos and the AI Cybersecurity Wake-Up Callbain.com
- 5OpenAI: Trusted Access for Cyber Defenseopenai.com
- 6Veracode 2025 GenAI Code Security Reportveracode.com
- 7arXiv:2508.14727: Static Analysis of LLM-Generated Java Codearxiv.org
- 8Nextdev: AI Coding Tools Pricing & 500-Developer Annual TCO Analysisjoinnextdev.com
- 9CrowdStrike Frontier AI Readiness & Resiliencecrowdstrike.com
- 10CrowdStrike Launches Project QuiltWorksir.crowdstrike.com
- 11CyberScoop: OpenAI Trusted Access for Cybercyberscoop.com
- 12TL;DR Sec #324 (Apiiro 322% privilege escalation & 10x findings summary)tldrsec.com