The GLM Sprint, Exploit Gap Collapse, and Grid Infrastructure

00:00:04 Open-weight models just had their practical coding moment ahead of Gemini. That's the headline from Nathan Lambert, a machine learning researcher who spent time at Anthropic and tracks this space closely. He noted that GLM-5.2 hit what he called a 'very practically useful in coding harness' milestone before Google even released its next model.

00:00:28 GLM-5.2 shipped. Within hours, Vercel published integration documentation for routing it through Claude Code via their AI Gateway platform. The gateway layer sits between your agent tooling and the model API — handling routing, observability, and in this case, model swapping without forcing you to rewrite your agentic loop.

00:00:51 Chris Tate at Vercel shared the docs showing how to configure Claude Code to send requests through the AI Gateway instead of calling Anthropic directly. You set ANTHROPIC_API_KEY to an empty string — which sounds weird but is necessary because Claude Code checks that variable first, and a non-empty value takes priority — then point it at your gateway endpoint.

00:01:16 Harrison Chase, who founded LangChain, shared a community guide showing how to build your own Claude Code-like agent using LangChain's Deep Agents framework. He flagged GLM-5.2's capability as the reason this matters now: if an open-weight model is good enough to slot into existing agentic tooling within hours, the integration race shifts from proprietary advantage to ecosystem velocity.

00:01:44 From a local-model angle, this lands on infrastructure — specifically the routing layer. Vercel AI Gateway doesn't just let you swap models; it lets you swap them in production with observability built in. A team can test GLM-5.2 against their Claude Code workflow without committing, then check real latency and quality numbers before deciding.

00:02:08 The integration cost of open weights just dropped significantly. The asymmetry here is worth tracking. Proprietary labs still control model access, but the tooling ecosystem — gateways, agent frameworks, coding harnesses — now moves in parallel to the release cycle rather than waiting for it.

00:02:29 Practical usefulness isn't measured by benchmark scores anymore. It's measured by how quickly someone wires a new model into an existing workflow. Vercel's docs also mention fast mode for Opus 4.6 and 4.7, which makes responses up to two and a half times faster at a higher per-token cost.

00:02:49 Same model, just prioritized differently. The infrastructure around these models is becoming the differentiator, not the weights themselves.

00:02:59 Anthropic published technical details about Claude Mythos Preview's cybersecurity capabilities today. The paper documents something notable: Nicholas Carlini, Keir Bradwell, and their team found that Mythos can identify and exploit zero-day vulnerabilities across every major operating system and web browser.

00:03:21 Their testing results are specific enough to be useful without being irresponsible, which matters because over ninety-nine percent of the discovered vulnerabilities hadn't been patched yet. The public examples include a web browser exploit combining four bugs with a complex JIT heap spray that escaped both renderer and OS sandboxes.

00:03:45 It also bypassed KASLR on Linux systems for local privilege escalation, and wrote a remote code execution exploit targeting FreeBSD's NFS server that granted full root access to unauthenticated users by splitting a twenty-gadget ROP chain across multiple packets.

00:04:04 The striking detail in the paper isn't what Mythos found — it's how they tested Opus 4.6 first. As Anthropic wrote last month, 'Opus 4.6 is currently far better at identifying and fixing vulnerabilities than at exploiting them.' Their internal evaluations showed near-zero success at autonomous exploit development.

00:04:26 Then they ran the same benchmark on Mythos. They took Firefox 147's JavaScript engine — all patches already applied in Firefox 148 — and asked both models to find and exploit vulnerabilities. Opus 4.6 turned the findings into working JavaScript shell exploits two times out of several hundred attempts.

00:04:48 Mythos achieved that same result one hundred eighty-one times, plus register control twenty-nine more. Across roughly seven thousand entry points in the OSS-Fuzz corpus with about a thousand open source repositories, Sonnet and Opus 4.6 hit basic crash tier between one hundred fifty and one seventy-five cases each.

00:05:10 Mythos hit five hundred ninety-five crashes at tiers one and two combined, plus ten full control flow hijacks on fully patched targets. Anthropic didn't explicitly train Mythos for these capabilities. They say the results emerged as a downstream consequence of improvements in code understanding, reasoning, and autonomy — the same improvements that make it better at patching bugs also make it better at exploiting them.

00:05:40 They're launching Project Glasswing to work with critical industry partners and open source developers before releasing Mythos-level capabilities more broadly. The paper frames this as 'a watershed moment for security,' which is bold phrasing from a company that tends toward measured language.

00:06:01 The local-model angle highlights the timing of the gap collapse. Last month, Opus 4.6 couldn't do this. A single generation later, it can. The paper's real tension isn't about what Mythos found — it's that there was a moment when the answer to whether frontier models could autonomously develop exploits was almost certainly no, and now the answer is definitely yes.

00:06:27 Anthropic says they expect powerful language models to benefit defenders more than attackers once defensive capabilities stabilize. They're right about the long term, but they also acknowledge that the transitional period could be tumultuous. It depends on whether labs are careful about how these capabilities are released.

00:06:50 Anthropic is rolling out identity verification for certain Claude capabilities starting July eighth. The support page says it's part of their routine platform integrity checks and applies to 'a few use cases.' It accepts government-issued photo IDs through a company called Persona.

00:07:10 That detail about Persona matters more than the policy itself. Persona is backed by Peter Thiel — not a neutral verification provider, but one with explicit investor ties to the AI ecosystem. If you fail verification, your account gets disabled with an appeal form as the only recourse.

00:07:29 The failure modes are familiar: blurry photos, expired documents, technical issues. Anthropic writes that they 'only use your verification data to confirm who you are and not for any other purposes.' That's a commitment to track closely, because the infrastructure layer is where commitments become architecture.

00:07:51 OpenAI already does identity verification for certain capabilities. The pattern across frontier labs is becoming clear: as model access moves from experimental tooling to production workflows, gating shifts from pricing tiers to identity tiers. The question isn't whether labs will do this — it's who controls the verification infrastructure and what happens when that control becomes a bottleneck.

00:08:19 I'm less interested in the policy details than in the structural implication. Identity verification at the API layer means someone else now controls part of your access pipeline. When Persona decides something has gone wrong, or when their service degrades, Anthropic's users feel it.

00:08:38 That's normal for any third-party dependency — but it becomes structural when identity is the gate.

00:08:45 Here's a concrete infrastructure story that doesn't involve speculation about nuclear fusion or decade-long grid buildouts. Second-life EV batteries are already powering data centers. The Forbes piece from Ken Silverstein profiles Re-Teck, which sits between retired EV battery supply and AI data center decommissioning services.

00:09:09 They receive battery packs from Tesla, Lucid, and BMW while also providing decommissioning for major AI operators including Microsoft. The connection is the key insight: they control both ends of the pipeline. Retired EV batteries retain roughly eighty percent of original capacity.

00:09:28 That's not enough for automotive use but it's well-suited for stationary storage. The concept is straightforward — charge during off-peak hours when electricity prices are lower, then discharge during peak demand periods created by AI workloads. Crusoe Energy already has a working deployment at their Sparks, Nevada campus.

00:09:51 Twelve megawatts of capacity and sixty-three megawatt-hours of storage, assembled from hundreds of repurposed EV battery packs. Ninety-nine point two percent operational availability since entering service. They're expanding from four to twenty-four modular data center units by early 2026.

00:10:12 The economics are moving in this direction too. New lithium-ion battery pack prices for stationary storage fell about forty-five percent last year, reaching roughly seventy dollars per kilowatt-hour. Repurposed packs from automakers can be significantly cheaper.

00:10:30 Re-Teck's CFO told Silverstein that nobody has a battery farm to supplement AI data center demand yet because it's too expensive — repurposed batteries are what breaks that barrier. The analogy they draw is with China's requirement for wind and solar facilities to pair with energy storage.

00:10:51 Instead of applying that on the generation side, Re-Teck applies it on the demand side: battery farms adjacent to data centers. This isn't the usual AI power narrative. Everyone talks about building new generation capacity — nuclear, solar arrays, geothermal — while the near-term constraint is actually buffering and storage.

00:11:14 The grid doesn't need more average power for most facilities; it needs peak shaving. Repurposed batteries are one path that's already working today. The piece also touches on supply chain dependency in a way that connects to broader infrastructure concerns. China controls roughly ninety percent of the global bismuth supply, eighty-five percent of rare-earth processing, and about half of U.S.

00:11:42 copper scrap ships overseas for processing. The battery reuse story intersects with these vulnerabilities: every pack repurposed domestically rather than exported represents a small step toward supply chain resilience. The EV revolution was largely promoted as an environmental response to fossil fuels.

00:12:03 Now those same batteries are helping power the infrastructure behind artificial intelligence.

00:12:10 Sonar's State of Code Developer Survey 2026 tracks the share of newly written code that was AI-generated or AI-assisted, and the trajectory is measurable. Six percent in 2023. Nineteen percent in 2024. Forty-two percent current for 2025. Fifty-five percent projected for 2026.

00:12:28 Sixty-five percent projected for 2027. The numbers come from a single methodology, so they track one specific slice of adoption rather than total output. The gap between 'AI-assisted' and 'AI-generated' matters — assistance includes completions, suggestions, and refactoring help in ways that change what the number represents.

00:12:51 And adoption data always carries selection effects: teams already comfortable with automation show higher rates by definition. What I find useful about this trajectory isn't the absolute percentage but the compression. We went from single digits to nearly half of all new code in three years.

00:13:11 The tools are shipping, teams are integrating them into existing workflows, and the gap between early adopter and everyone else is collapsing faster than previous developer tooling migrations. The open-weight momentum we opened with — GLM-5.2 slotting into Claude Code via gateway in hours, community guides from LangChain's founder showing you can build your own version — connects to this trajectory on the supply side.

00:13:41 More model options at better price points mean more teams can experiment without committing to a single vendor's pricing. Anthropic's Mythos paper and the Claude ID verification rollout connect on the governance side. As access expands and capabilities increase, the questions about who controls what shift from technical decisions to structural ones.

00:14:05 The battery story connects everything: infrastructure constraints that affect both model availability and code tooling adoption timelines. Everyone's racing toward similar integration points — gateway layers, verification infrastructure, energy buffering — and the near-term winners will be the teams that treat those as engineering problems rather than market positioning.

00:14:31 Infrastructure is outpacing capabilities again. — Seln Oriax.