NEWS

Claude text watermarking: Claude models released after August 2, 2026 embed a watermark in generated text and C2PA provenance in files. What is known so far.

Cactus open-sourced Needle 2, a 45M-parameter agentic LLM in a single 14MB binary that runs a full tool-calling session in 28MB of RAM. 500 tok/s on a Raspberry Pi 5, ESP32-S3 class parts, Apache 2.0. Here is what the benchmarks actually show.

Cloudflare's H1 2026 DDoS report: attacks above 1 Tbps grew 519% quarter over quarter as DNS floods became the leading network-layer vector.

GitHub Copilot SDK for Java adds @CopilotTool annotations, virtual threads, Spring and Jakarta EE composition, and BYOK against any OpenAI-compatible endpoint.

Encrypted chain-of-thought from Anthropic, OpenAI and Google APIs can be replayed into weaker models and read verbatim. How the attack works and what leaked.

A new arXiv paper probes six widely used open-source agent frameworks and finds the barrier semantics of approval gates, cancellation, and timeouts hold on none of them. A sibling branch can execute while the user is rejecting another one, and replay can double-execute. The fix is a verified external gate called SoundGate.

Vercel Sandbox egress control polices outbound traffic on the host with domain policies, CIDR rules, credential injection and deny-all by default.

AgentChaos injects server errors, truncated responses and corrupted tool calls into live agents. Pass@1 drops up to 50 points: robustness is a systems problem.

LLM distillation on one GPU: cache the teacher's top-100 logits offline and use a fused chunked KL loss to cut a GPT-OSS 20B run from four nodes to one.

LivePlan monitors SWE-agent runs and calls an advisor LLM only on drift, raising resolution rates up to 15.2 points for an extra $0.08 per instance.

Muse Glimmer 30B is Meta's Apache 2.0 open-weight multimodal agent model that runs in 24GB. Benchmarks, hardware reality and what the numbers actually say.

GPT-5.6-Cyber is OpenAI's gated security model, available only through Daybreak Red. The benchmark numbers, the Chrome V8 CVE it found, and who gets access.

Claude Code v2.1.224 lets one running session message another over a first-party channel - plain text, permission-aware, with approval dialogs when bypass-mode sessions talk to each other. Here is what ships, how delivery and inbound controls work, and where the feature stops.

A Claude-built night-sky site cloned an open source project's name, features and even a fixed bug. What it says about memorization and AI app verification.

Copilot code review's Lite and Balanced effort levels are generally available. Here is what each level does, what it costs in AI credits and Actions minutes, and how to set org-level defaults so review depth matches PR risk.

GitHub Copilot's impact dashboard now models ROI in dollars from AI credit use, PR output and salary. What it says about agent-first vs passive adoption.

GitHub Apps can now be installed at the enterprise level, opening enterprise management to third-party integrators, with a security boundary on top permissions.

Grok Imagine Image 2.0 is xAI's typography-aware image model, ranked second on text-to-image and editing leaderboards and callable via Vercel's AI Gateway.

Hermes Agent now supports Vercel AI Gateway and Vercel Sandbox. What shipped, how to switch, and why agent infrastructure is going bring-your-own.

Vercel Skill Packs bundle agent skills into one installable, unlisted pack on skills.sh. How packs work, what they mean for teams, and the trust questions.