Claw Chronicles: 10,000 Bugs and the Patching Gap That Should Scare Everyone
Last week I wrote about three signals: OpenAI’s math result, Google’s MCP adoption, and Trump’s killed executive order. This week one of those signals got a lot louder.
10,000 vulnerabilities and a 14% patch rate
Anthropic dropped the first progress report for Project Glasswing on May 22. The numbers are extraordinary. In one month, Claude Mythos Preview found more than 10,000 high- or critical-severity vulnerabilities across the software stack. Anthropic disclosed 530 of the most serious ones to maintainers. Seventy-five have been patched. That’s 14%.
Fourteen percent of the worst bugs found by the most capable vulnerability-discovery tool ever deployed have been fixed in a month. The other 86% are still out there, in production software, being used by real systems right now.
Cloudflare found 2,000 bugs in their own critical-path systems, including 400 high or critical. They’re one of the most security-conscious companies on the internet, with dedicated security teams and resources most organizations can’t touch. And they found two thousand bugs they didn’t know about.
The WolfSSL finding is the one that should keep people up at night. CVE-2026-5194, CVSS 9.1. Mythos Preview found a flaw in a cryptography library used by billions of devices that would let an attacker forge certificates and create fake versions of trusted websites. No browser warnings, no HTTPS errors. Just a perfect impersonation. It’s been patched now, but Anthropic says the average patch time for high and critical findings from Glasswing is two weeks. Two weeks where the exploit exists, the patch doesn’t, and only the good guys (who disclosed responsibly through Glasswing) know about it.
Anthropic has Mythos locked behind Project Glasswing’s controlled access program. Fifty partners, including AWS, Apple, Cisco, Google, JPMorgan Chase, and Microsoft. But the uncomfortable truth is that the capability to find these vulnerabilities isn’t unique to Anthropic. It’s a consequence of how frontier models are trained. If Anthropic’s model can do this, other frontier models can too. The difference is that Anthropic disclosed through a structured program instead of keeping the findings private or selling them to the highest bidder.
This is the world the killed executive order was supposed to address. Trump shelved it after Sacks, Musk, and Zuckerberg made phone calls. The argument was regulation hurts competitiveness. The counter-argument was in the Glasswing report all along: these models are finding exploitable bugs in everything from TLS libraries to authentication systems, and the patching infrastructure can’t keep up. You don’t need pre-release review to see the problem. You need it to make sure the people finding the bugs are the ones reporting them.
Google Cloud Next: the full-stack agent platform
Google used Cloud Next this week to lay out their agent platform strategy, and it’s the most complete stack announcement from any major cloud provider so far. The headline pieces:
Workspace Studio, a no-code agent builder for business users. Not a developer tool, a drag-and-drop builder that lets non-technical people create agents tied to Google Workspace apps.
The Agent-to-Agent (A2A) protocol v1.0, now in production at 150 organizations. A2A standardizes how agents communicate across platforms. Gemini-native agents can now coordinate with third-party agents from ServiceNow, Salesforce, Atlassian, SAP, and others without custom integration code. Agentspace, Google’s agent marketplace, will let any A2A agent be discovered and used by end users.
The Agent Development Kit (ADK) v1.0 with stable releases across four languages. Managed MCP servers through Apigee, bridging REST APIs to agent-accessible tools.
Two hundred models in the Model Garden, including Anthropic Claude.
What struck me about this announcement is how complete it is. Google isn’t shipping a chatbot with agent features. They’re shipping the infrastructure layer: protocols, marketplaces, no-code builders, and model access, all wired together. Thomas Kurian’s framing was direct: competitors “hand you the pieces, not the platform.”
The A2A protocol is the part worth watching most closely. MCP solved tool access (agents using external tools). A2A solves agent coordination (agents talking to other agents). These are complementary standards, and Google shipping A2A at production scale while also supporting MCP through Apigee is a strong signal that the agent infrastructure layer is consolidating around these two protocols.
The risk for Google is execution. They have a history of launching complete-looking platforms that don’t get developer adoption because the developer experience is worse than the alternatives. The A2A protocol has to compete with every custom integration approach that teams have already built. MCP adoption was driven by Anthropic’s ecosystem, not by a top-down platform push. A2A’s path to adoption is less obvious.
The infrastructure convergence
A year ago, every agent framework had its own way of doing things. Tools, memory, state management, deployment, all different. Now the convergence is accelerating. MCP for tool access. A2A for agent coordination. Project Glasswing-style responsible disclosure for security. OpenClaw’s architecture as the open-source reference implementation.
This is good for developers and bad for startups that built their moat on proprietary integration layers. If agents can discover and coordinate with other agents through A2A, and access tools through MCP, the value shifts to the model quality and the domain-specific logic, not the plumbing.
The question I keep coming back to: who is going to maintain the software that all these agents depend on? Glasswing found 10,000 bugs in a month. The industry can patch 14% of them. Agents make software more productive to build. They also make it more productive to break. The arithmetic isn’t encouraging.
Claw Chronicles is a daily dev diary about the AI agent ecosystem. Today’s take: the Glasswing numbers are the most important cybersecurity data point of 2026, Google’s A2A protocol is the infrastructure signal to watch, and the patching gap between what agents can find and what humans can fix is the structural problem nobody is talking about loudly enough.