AI #174: You're It
Zvi's AI Roundups · Zvi Mowshowitz · 2026-06-25
Zvi Mowshowitz's weekly AI roundup covers Anthropic's Claude Tag Slack integration, White House negotiations with Anthropic over a jailbreak severity framework, GLM-5.2 as a potential open-source agent breakthrough, Alex Bores's narrow loss in the NY-12 congressional race, Google's AI control roadmap, and growing concern about under-secured AI coding agent deployments.
Appears in
Extraction
Topics: ai-agentsai-policyai-safetyopen-source-modelsai-governanceai-politics
Claims
- Anthropic's Claude Tag integrates Claude into Slack as a persistent, asynchronous, memory-bearing agent, with Anthropic reporting it now generates 65% of their product team's code.
- The White House and Anthropic are jointly developing a formal technical framework to quantify jailbreak severity rather than treating incidents as binary pass/fail events.
- GLM-5.2 represents a potential 'DeepSeek moment' for open-source AI agents, making frontier agentic capabilities available in open weights for the first time.
- Alex Bores lost the NY-12 congressional race by 4%, with the winner Micah Lasher supporting the RAISE Act, making the outcome ambiguous or a net negative for the AI-PAC that opposed Bores.
- Google's AI Control Roadmap v0.1 outlines defense-in-depth mitigations against misaligned AI but Zvi argues it underestimates the threat model and that CoT monitoring will become unreliable before the roadmap's trigger thresholds are reached.
- AI coding agents are being routinely deployed with broad organizational access—including internet connectivity, internal secrets, and autonomous overnight execution—without adequate security auditing.
Key quotes
Andrej Karpathy (Anthropic): This is a new paradigm for interacting with Claude that is significantly more 'inline' with all the other human activity org-wide... Claude basically joins the team in a seamless way - you can talk to it as you would talk to a person and it can help with a very large variety of workloads.
Marius Hobbhahn: 'So you're telling me that you have AI agents running on your organization's computers that: Have access to the internet... Often run autonomously for hours without human oversight... And you don't even analyze whether the agents have done something malicious after they have run, i.e. you're just flying blind?'
roon (OpenAI): the world must skate between antichrist and armageddon and it looks increasingly difficult.