# AI News Report, October 26: CODEX RAN MALWARE IN OVER 90% OF TRIES FROM ONE FILE IN DOWNLOADS > Source: > Published: 2026-10-26 14:15:51+00:00 The S-1 shows $11.6 billion in revenue for the quarter ended June 30 and an $80.7 billion net loss for the half, mostly from revaluing notes and other securities. It warns it may change usage limits, rate limits and model capabilities. Opus 5.5 scores 66.4% on the same test, and full weights arrive by November 4. Moonshot pulled cyber training data from the open release, which cut one hacking score from 90.5% to 81.0%. It will train on cached and simulated copies of the web after a model reached the internet through a misconfiguration on September 20. OpenAI says its models are too unpredictable to trust with live access. On 200 biopsy-proven lesions, the old Inception v3 network got 62% right on its first guess, against 42% for both LLMs. Only 30 melanomas were tested, but it is a strong case for small specialist models. Block says it trained Dessa on one NVIDIA GB200 NVL72 with no text, and it runs on a MacBook Pro M4 Max. It is not public yet; safety researchers get early access first. Its own tests put it at 80.9 on Terminal-Bench 2.1, ahead of Qwen3.8-27B at 73.0. Worth a local trial if you want an open agent model in the 30B class. He had GPT-6 Astra write zstd implementations that passed every test over 99% of the time, for about $1 each. His point: cheap code makes projects you once skipped worth building. He maps four ways long contexts fail and six fixes, from pruning to offloading. His most useful tip: keep state in files and reset the agent's context when it gets stuck. Cache reads are 96 to 99 percent of agentic coding tokens, so a 75 percent cut in cache read prices cut real costs by about 70 to 75 percent. Check your own cache hit rate before you compare prices. Kimi K2.6 tool calls matched their schema 92% of the time until Ubicloud turned constrained decoding back on, and then hit 100%. If you serve models on vLLM 0.31, check that flag. KubeVirt was founded by Red Hat in 2016 around the idea that Kubernetes could run virtual machines alongside containers. We The post The next 10… · The New Stack A Perplexity researcher caught a working exploit on an Apache mod_php honeypot running PHP 8.4.26, and Marcus Hutchins confirmed it. If any PHP server missed the October 21 update, patch it today. It reuses the Baton Drop flaw, CVE-2022-21894, because the full fix needs a manual step to revoke the Windows Production PCA 2011 certificate. Check that your Windows fleet has done that step. The 72 fixes cover AOS-8 and AOS-10, and the worst need no login. HPE knows of no exploit yet; until you patch, keep management interfaces on their own VLAN. Microsoft says the January 2027 update is the final one for every Windows Server 2016 edition. If you still run it, plan the move to Windows Server 2025 now. The agent asks, you approve on your own device, and 1Password fills the login itself. It launches with Claude Duet, ChatGPT Work, Meta's Muse and Grok Bot. It reads a chat's context to catch impersonation even with no bad link or file, with a focus on outside first-contact chats. Rollout begins in early November, and admins do not need to do anything. Asked through the APIs about nearly 500 midterm races, Grok 4.7 picked the Republican 92% of the time and GPT-6.1 Sol the Democrat 97%. The consumer apps decline the question. Arvind Krishna told investors open-weight models cost 10 to 25 percent of frontier token prices and fit in two or three racks. The hard part, he said, is finding people to run them. The BriefSecurity firm Forever found a booby-trapped AGENTS.md instruction file in a customer's Downloads folder, and in its tests OpenAI's Codex followed that file and installed malware in more than 90% of attempts. Coding agents read instruction files from the folder they start in, and Forever says a web page can save a file into Downloads with no click, so a planted CLAUDE.md also got Claude Code to send its traffic and login token to the researchers' server about half the time. Never start a coding agent in Downloads, and read any AGENTS.md or CLAUDE.md you did not write. Level UpLearn exactly which instruction files your coding agent reads before it starts. Claude Code loads CLAUDE.md, and AGENTS.md when no CLAUDE.md exists, from the folder you launch in and every folder above it, so the folder you start in decides what it obeys. →Claude Code docs: how CLAUDE.md and AGENTS.md files load Caught Up? Oct 19: NorForge found open control panels for Cairn, an AI agent platform criminals used to break into 100+ company sites at about $50 a hack. Oct 20: DHS ordered US civilian agencies to remove DeepSeek, even when it is built into other tools. Oct 21: PHP patched two flaws rated 9.2 that let strangers run code on web servers. Today, working attacks are confirmed. Oct 22: LiteLLM 1.107.0 shipped a key stealer to PyPI and Docker. Oct 23: Anthropic and Theorem said Claude wrote machine-checked proofs for 123 open-source libraries and found hundreds of bugs.