Anthropic's agent thesis, open-weight bug hunters, and Vite's team joins Cloudflare

When agents write most of the code

Anthropic's Institute published an essay arguing that AI development is entering a recursive loop in which coding and research agents accelerate their own improvement, so productivity gains compound instead of arriving as one-off boosts. Anthropic also offered an internal number. VentureBeat reports the company says 80% of its newly authored production code now comes from Claude, which the piece presents as a preview for enterprises, with consequences for review, code ownership and hiring. Spotify reaches a related conclusion from its own experience: with tools like Claude Code in use, coordination, review and system design have become the limiting factors, not raw code production.

A leaked thread claims Anthropic has made a model called claude-oceanus-v1-p available to external red teams. Details are thin. The red-team-first rollout resembles how Anthropic has previewed earlier model generations, so teams that plan migrations around its release cadence may want to watch for a formal announcement.

Bug finding depends on the harness

AI-assisted vulnerability research produced a concrete result in mature infrastructure this week. An AI security tool uncovered a two-year-old Redis bug that lets attackers take over servers, a flaw that conventional review and fuzzing had missed. Two posts examine how much of that kind of capability belongs to the model itself. One reproduces a Mythos-attributed FreeBSD find using local open-weight models, and the other looks at the harness and post-training techniques that close the open-weight gap. Both argue that tooling, retries and structured search matter as much as raw model quality, which is relevant for teams building internal vulnerability-scanning pipelines.

For coding work, a separate assessment of local LLMs for agentic coding compares them with cloud frontier models on cost, latency and reliability. Its author finds self-hosted models closing the gap on narrower agentic tasks but still lagging on complex multi-step reasoning.

Sandboxes, audits and disposable clusters

Token Security's Zapocalypse writeup shows how two modest primitives can combine into a platform-level risk. An unrestricted os.system call allowed a Python sandbox escape, and regex scraping recovered orphaned STS tokens, which together could have let an attacker hijack automation workflows across Zapier. It is a useful case study for anyone shipping sandboxed code execution in a SaaS product.

On the Kubernetes side, CNCF published results from the first formal security audit of Inspektor Gadget, the eBPF-based debugging and observability toolkit, giving production users a clearer view of its security posture. Pulumi describes an EKS Environment Factory that pairs Pulumi with vCluster virtual clusters, producing cheap, isolated, ephemeral environments on a shared EKS cluster for per-PR or per-team testing.

Vite moves in, Elixir adds types

VoidZero, the company behind Vite and the Rolldown bundler, is joining Cloudflare, which extends Cloudflare's reach from edge compute down to the build step. It is not yet clear how these tools will evolve inside a larger platform rather than as independent open-source projects. Elixir, meanwhile, released v1.20 as a gradually typed language, and the release notes describe a type system that teams can adopt incrementally without rewriting existing code.