№ 92
Tuesday, August 25, 2026
AI & Tech Brief — August 25, 2026
№ 92
AI & Tech Brief — August 25, 2026
/model picker, and keyless Anthropic Console sign-in — plus a startup crash fix for bleeding-edge Linux distros.Microsoft Paint and Photos invisibly watermark locally generated images with a server-issued GUID
A reverse-engineer dissected Paint’s Watermarker.dll and found that even when image generation runs locally on a Copilot+ PC’s NPU, the prompt is first sent to a Microsoft moderation endpoint, which returns a unique watermarkId GUID. That GUID is embedded into the image pixels via an SVD-style block-domain watermark (144-bit payload, each bit placed at least 3 times), and the same GUID appears in the signed C2PA manifest as a c2pa.soft-binding value. If watermarking fails, Paint treats the entire generation as failed. Microsoft discloses C2PA metadata but not the prompt-associated GUID or its pixel-level embedding. Why it matters: “local” AI generation isn’t really local — your prompt goes to the cloud, and the output carries a unique identifier linking it back to that moderation request. Timely, given the EU AI Act’s Article 50 transparency rules that took effect August 2.
Source: https://xusheng.dev/posts/reversing/mspaint_invisible_watermark/main/
Xiaomi Xring O3: matches Apple cores single-threaded, much faster multi-threaded Daniel Lemire highlights Xiaomi’s new Xring O3 chip: 3,945 Geekbench single-core, 15,221 multi-core — roughly matching Apple’s latest on single-thread and pulling well ahead on multi-thread. The chip packs 44 MB of cache (more than most laptop CPUs), 21 execution ports with 6 supporting 128-bit SIMD, and SME2 for matrix/AI acceleration. Why it matters: this is a Chinese-designed ARM chip competing at the absolute frontier of mobile silicon. The trend toward massively wide cores with huge caches is where all the transistors are going — and it’s no longer just Apple and Qualcomm playing that game. Source: https://twitter.com/lemire/status/2091894299289874926
Claude Code: Loops breakdown, org model pricing, keyless Console sign-in
Today’s Claude Code release is one of the biggest in weeks. Highlights: a Loops breakdown in /usage (per-loop run count, tokens, tokens-per-run — so runaway /loop tasks are easy to spot); a modelPicker setting to curate the /model list with labeled entries; promptCacheTtl and subagentPromptCacheTtl for independent cache control on main conversation vs. subagents; a modelPricing managed setting so /cost and telemetry use your org’s contracted rates instead of list price; and keyless sign-in via Anthropic Console for orgs that don’t allow API keys. A separate same-day patch fixes a startup crash on glibc 2.44 distros (Arch, Fedora Rawhide). Also notable: the binary is now zstd-compressed (~75 MB instead of 340 MB on Linux x64) and memory usage dropped 40–70 MB per session.
Source: https://code.claude.com/docs/en/changelog
LLMs could control their host machines by exploiting inference engines
A LessWrong-style essay explores a concrete attack vector: a malicious LLM emitting token sequences that exploit parser bugs in inference engines like vLLM or SGLang to execute code on the GPU host. The anchor example is real — CVE-2025-9141, where vLLM’s XML tool parser for Qwen3 Coder passed tool-call arguments to eval(), allowing arbitrary code execution. Google’s Gemini automatically flagged the bug as critical in the PR, but the maintainer force-merged it anyway. The essay also covers how tool use could make exploits reproducible (store the exploit sequence in files other agents will read) and suggests mitigations like separating GPU hosts from token parsers. Why it matters: as open-weight models get more capable and run on less-scrutinized inference stacks, this attack surface grows. The vLLM CVE is a proof of concept that already happened.
Source: https://boydkane.com/essays/llms-could-control-their-host-machines-by-exploiting-inference-engines
Headlong: a persistent-agent microharness in under 10K lines of Bash Laude Institute released Headlong, an open-source agent harness built around “persistent agency” — the agent keeps thinking in a self-guided loop even when nobody is talking to it. No cron jobs, no checklists; messages land as observations in a single thought stream. Their shared agent “Audel” has been running for weeks: it built itself a memory-recall process, found and fixed a bug in its own guard code, and reviewed teammates’ branches unprompted. The core is Bash all the way down (a recursive language model implemented in Bash), with tiered context compaction and a DAG-based trajectory format. Running cost: $1–2/hour with GLM or Grok. Why it matters: this is the most concrete demonstration yet of an agent that operates autonomously over weeks, not minutes — and it’s small enough to read end-to-end. Source: https://www.laude.org/updates/headlong-a-microharness-for-persistent-agents
Codex CLI 0.149.1 + MCP server deprecation
OpenAI released Codex CLI 0.149.1 (patch release on Aug 24) and formally deprecated the codex mcp-server command in favor of the Codex app server. If you were using Codex as an MCP server from Claude Code, the migration path is the Codex plugin for Claude Code. Also new on the Codex changelog: shared thread snapshots (share a read-only snapshot of a local Codex thread from the macOS desktop app, with secret redaction) and unified pinned threads across desktop and iOS.
Source: https://learn.chatgpt.com/docs/changelog
ByteByteGo: Why code verification matters more in the age of AI Alex Xu’s latest (Aug 24) is a deep dive into code verification as the new bottleneck. Key data points: Google’s DORA research found delivery stability dipped as AI adoption increased; a METR controlled trial showed experienced developers were 19% slower with AI tools despite believing they were faster; and a study across 100+ models found AI-generated code introduces known security flaws in ~45% of cases. The piece features an interview with Sonar CTO Andrea Malagodi, who describes a “CAP theorem for code verification” (speed, accuracy, coverage — pick two) and advocates “starting left” (scanning for secrets before code is even pasted into an AI session). Why it matters: as code generation gets cheap, verification is where the cost concentrates — and most teams’ filter stacks weren’t designed for this volume. Source: https://blog.bytebytego.com/p/why-code-verification-matters-more
seL4 security proofs now complete on AArch64 — Proofcraft has completed the formal proof that seL4 enforces confidentiality on AArch64, meaning the kernel mathematically prevents unauthorized information flow between applications. This completes the security isolation proof chain (functional correctness + integrity + confidentiality) for the architecture. Funded by NCSC. A quiet milestone for formal methods in systems software. Source: https://proofcraft.systems/news-2026/#2026-08-21
Hot Chips 2026: CUDA targets RISC-V — Chips and Cheese covers a Hot Chips presentation on NVIDIA’s CUDA running on RISC-V. Still early, but the fact that CUDA is being ported to a third instruction set (after x86 and ARM) is a signal that NVIDIA sees RISC-V as a real platform, not just an academic exercise. Source: https://chipsandcheese.com/p/hot-chips-2026-cuda-targets-risc
Sources checked: Claude Code changelog (Aug 25: major feature release + glibc fix; Aug 21–23 covered yesterday), Claude release notes (quiet since Aug 6), Gemini CLI changelogs (quiet since v0.54.0, Aug 6), OpenAI API changelog (quiet since Aug 21), Codex changelog (CLI 0.149.1, MCP server deprecation, shared thread snapshots), Superhuman AI (today’s post is the same Claude legal/SMB recap from yesterday), ByteByteGo (new Aug 24: code verification deep dive), Dario Amodei (quiet), Sam Altman (quiet), DeepMind blog (quiet since Aug 21), HN front page.