🤚 The Open-Palm Repository

Remember Project Glasswing, the initiative where Anthropic pointed its most powerful model at open-source software and it found over 10,000 high- and critical-severity vulnerabilities in its first few weeks? The one we covered in May when the security community collectively said “cool, terrifying, thank you”?

Well, Anthropic has now open-sourced the entire framework. The Defending Code Reference Harness dropped on GitHub this week, and it’s exactly what it sounds like: a seven-stage autonomous pipeline that discovers vulnerabilities, verifies them, deduplicates findings, writes exploitability reports, and generates patches — all without a human touching the keyboard.

The repo has already accumulated 1,300 GitHub stars and 101 forks. It’s written primarily in Python (92.7%), runs inside Docker containers with gVisor sandbox isolation, and works with Claude via direct API, AWS Bedrock, Google Vertex, or Azure. The framework currently targets C/C++ memory vulnerabilities using AddressSanitizer, but is designed to be portable to other languages and vulnerability classes.

👐 The Two-Handed Autopsy of the Pipeline

The seven-stage pipeline reads like someone turned a penetration testing team into a flowchart and then replaced every person in the flowchart with an API call:

• Build — Compile the target with instrumentation (ASAN for C/C++)
• Recon — Partition the codebase into distinct attack surfaces
• Find — Parallel agents craft inputs and trigger crashes
• Verify — An independent grader agent reproduces each crash
• Dedupe — A judge agent removes duplicates
• Report — Structured exploitability analysis
• Patch — Generate and validate fixes

Notice how there are three different AI agents checking each other’s work: one finds the bug, one verifies it independently, and one judges whether it’s a duplicate. This is peer review, except the peers are all the same entity wearing different hats. It’s like a law firm where every partner is the same person but they still bill separately.

The interactive skills include /threat-model, /vuln-scan, /triage, and /patch — Claude Code slash commands that let security teams run individual stages manually. Anthropic’s recommended ramp-up: threat model and static scan on Day 1, run the full pipeline on Day 2, customize for your targets by Day 5, and scale autonomous scanning by Week 2. That’s a faster onboarding than most companies give their human security hires.

🌿 The Gentle Awakening

There’s a peculiar irony in open-sourcing a tool that automates the work of the very people you’re giving it to. Anthropic is essentially handing the security community a machine that does their job — and the security community is thrilled about it, because unlike most automation stories, the alternative here wasn’t “a human does it carefully.” The alternative was “nobody does it at all because there aren’t enough humans.”

The cybersecurity talent shortage has been a running joke for a decade — somewhere between 3.4 and 4 million unfilled positions globally, depending on which report you read and how existential you’re feeling. When Glasswing found 10,000 vulnerabilities in weeks, it wasn’t replacing security researchers. It was doing work that was never going to happen otherwise, because the queue of unaudited code is effectively infinite.

This is the rare case where “AI replaces workers” and “AI creates more work for workers” are both true simultaneously. The tool finds the bugs. Humans still need to decide what to do about them, how to prioritize remediation, and whether the AI-generated patch is trustworthy. The bottleneck shifts, as it always does — from finding vulnerabilities to caring about vulnerabilities.

👑 The Gold-Leaf Security Forecast

Anthropic also offers Claude Security as a managed, hosted product — making the open-source release a strategic move as much as a philanthropic one. Give away the framework, sell the infrastructure. It’s the MySQL playbook, except the product finds zero-days instead of storing your blog posts.

The repo’s README states it is “not actively maintained” and is “not accepting contributions,” which is corporate for “we built this, it works, here’s the blueprint, please don’t open a pull request asking us to add COBOL support.” The real investment is in the managed service, where Anthropic has been partnering with security teams to deploy Mythos Preview on actual codebases at scale.

For the rest of us, the message is clear: the era of AI-powered vulnerability discovery is no longer a research preview or a press release. It’s a GitHub repo with a Dockerfile and a seven-step pipeline. The barrier to entry just dropped from “hire Anthropic” to “clone the repo.” Whether you find that reassuring or terrifying depends entirely on which side of the vulnerability you’re standing on.

“The AI found the vulnerability, the AI verified the vulnerability, the AI deduplicated the vulnerability, and the AI wrote the patch. The human’s job is to click ‘merge’ and update the dependency tree. We have automated the entire security lifecycle except for the part where someone takes credit on LinkedIn.” — The Slap of Wisdom Vulnerability Disclosure Team, running the pipeline on its own WordPress installation and hoping for the best