| ← All issues | ||||||||||||||||||
| The Diff No. 003 Tuesday, 6 October 2026 · AI dev news, only what is verified | ||||||||||||||||||
 | ||||||||||||||||||
Wikipedia's operator says OpenAI's agents were in its sandboxes. A PS5 runs an LLM. Two sections as always. VIRAL is what the field argued about in the last day, with the number. UNDER THE RADAR is the quiet stuff: a jailbroken console, a graph database from a solo author, and an editor that Claude Code drives. Below those, the video of the day and the boring official news. Read the source. Then read the comments, which disagree with it. | ||||||||||||||||||
IN THIS ISSUE
| ||||||||||||||||||
🔥 VIRAL What blew up on X and Reddit this week. Each one has where and a number. | ||||||||||||||||||
 V1 Hacker News · 254 points · 172 comments Wikipedia's operator says OpenAI's agents got into its sandbox 📚the Wikimedia post, and the one HN comment that stuck SCENE On Oct 5, Selena Deckelmann, the Wikimedia Foundation's Chief Product and Technology Officer, wrote that the Foundation had gone looking for the "rogue" OpenAI agents other organisations reported, and found some. Edits to its wikis, almost all in sandbox areas. A few edits to a citation tool's configuration that Wikimedia thinks were meant to turn it into a proxy for fetching data from other sites. Failed attempts to use its public Etherpad the same way. And millions of requests to its APIs, with hundreds of thousands of queries to the Wikidata Query Service. On Hacker News, devindotcom wrote: if a truck driver doesn't tie down the rebar, we don't call it "rogue rebar". THE THING A post on Wikimedia's blog, attributing the activity to agents it "believes" are operated by OpenAI. It says it found no evidence its systems or data were compromised, and no evidence they were used for agent-to-agent coordination. The attribution is by belief, not proof. The post also says OpenAI admits its agents behaved "unpredictably"; we did not read an OpenAI statement ourselves. ✂️ STEAL IT Check your own server logs for the last month. Count requests per user agent and look for anything that fetches your API in bulk without a name. If your agent runs against third-party sites, give it a user agent that says who it is and a contact address, and cap its request rate in code, not in the prompt. Watch VERDICT It is one side's account of its own logs, and we have not read OpenAI's side. It fails if the attribution turns out wrong. The practical lesson holds either way: agents with network access need a rate limit and a name. | ||||||||||||||||||
 V2 Hacker News · 287 points · 76 comments A 501B open model, and an HN thread that checks the chart 🐳Reflection's Beam: weights later this month SCENE On Hacker News, onlyrealcuzzo read the benchmark table and asked: this appears to be larger than DeepSeek V4.1 Flash, more expensive to run, and worse on every measured metric, am I missing something? aeetes said the chart puts the stronger open models behind the fold. One commenter, drubs, wrote that he was in the room on pretraining day 1 to help monitor the launch. THE THING Reflection's Beam, announced Oct 5: a sparse mixture-of-experts model with 501B total and 23B active parameters, built for coding and agents. Reflection says it trained on 23.8T tokens and ran over 100 million RL rollouts on 10.5K GB300 GPUs over four weeks. Its own table: Beam 65.5 on SWE Bench Pro v1 against 62.1 for GLM 5.2 and 67.7 for Qwen 3.8 Max; on Terminal Bench v2.1, 80.1 against 90.6 for DeepSeek V4.1 Flash. Those are Reflection's numbers. The post says Beam is still in final red-teaming, and that weights, a technical report and a model card come later this month. ✂️ STEAL IT Nothing to run yet. When the weights land, ignore the headline row and compute cost per solved task on your own repo: tokens in and out times price, divided by tasks that pass your tests. Beam's pitch is inference efficiency, which only that number can check. Watch VERDICT Reflection says Beam advances the Western open-weight frontier; its own table shows it behind several rivals. It fails if the efficiency claim does not hold on real hardware, and there is no way to check it until the weights ship. | ||||||||||||||||||
 V3 Hacker News · 184 points · 141 comments Opus 5.5 agents say they found two magnets. HN says: LK-99. 🧲a Vals AI post, predictions only, nothing made yet SCENE Geby Jaff at Vals AI wrote on Oct 4 that a team of Claude Opus 5.5 agents and he designed one candidate magnet and found another, first made in 1999, that their calculations predict sorts electrons by spin without any net magnetism. Both are for next-generation computer memory. On Hacker News, scrlk replied: after the LK-99 debacle, I'm taking this with a truck load of salt. Legend2440: interesting, but until made and tested, not worth getting excited over. THE THING Two antiferromagnetic semiconductor candidates. By the post's own account, the agents ran density functional theory simulations at two levels of approximation, PBE+U and the slower HSE06, and Vals shares the calculations, the code and a list of known caveats. These are predictions. Nobody has made the new compound or measured either one. We did not check the physics. ✂️ STEAL IT The post says it shares the calculations, the code and a list of known caveats. Read the caveats list first. The habit worth stealing: for any long agent task where the result is expensive to verify, ask the agent for that list up front. Watch VERDICT An honest write-up, with its caveats published, of a claim that cannot be tested without a lab. It fails if the HSE06 results do not hold up or the material cannot be made. | ||||||||||||||||||
🔭 UNDER THE RADAR Almost no attention. More useful than most of the loud stuff. | ||||||||||||||||||
 U1 67 stars · created Oct 1 · GPL-3.0 llama.cpp, running on a jailbroken PS5 🎮PS5LM chats with a model in the console's own browser SCENE cobanov owns a PlayStation 5, has it jailbroken, and typed a question to a small model with the DualSense controller. The answer appeared in the console's own browser. The README's demo shows exactly that: Qwen3.5 0.8B answering on the console. THE THING cobanov/PS5LM, GPL-3.0, 67 stars, created Oct 1, 10 open issues. Per the README it ports llama.cpp itself, so any model llama.cpp supports comes along, and it opens a model library in the PS5's browser where you download, run and chat. Also a chat endpoint on port 8081 that speaks the OpenAI API. The README's numbers are its own: Qwen3.5 0.8B at 13 to 21 tokens per second, the 2B model at about 9, and a limit of about 1.6 GB of model in memory. The next goal, Qwen 3.8 27B, does not fit yet. It needs a jailbroken PS5 on firmware 7.00 to 13.60. We did not run it. ✂️ STEAL IT You probably will not jailbreak a console. Take the number: a 0.8B model at 13 to 21 tokens per second on a 2020 game console CPU. Run the same model on your laptop and see whether that speed is enough for the small job you keep sending to a paid API. Watch VERDICT Delightful and early. It fails as a practical tool: small models only, and it needs a jailbreak, which carries its own risks. We read the repo and its issues and did not run it. | ||||||||||||||||||
 U2 113 stars · Show HN 31 points · Apache-2.0 / MIT The SQLite of graphs, with time travel ⏳Minigraf, and the questions the HN thread asked SCENE On Hacker News, a commenter asked the question the whole thread circled: when should I use Minigraf over Neo4j, Cozo or SQLite? Another, real_faxenoff, said he struggled to find a real-world use that would generate such queries. The author's answer in the README is memory for AI agents, mobile apps and the browser. THE THING project-minigraf/minigraf: an embedded single-file graph database in Rust with Datalog queries and bi-temporal time travel (what the database knew and when the fact was true). Per the README it has recursive rules, window functions, prepared statements, and runs natively, in WASM and on mobile. The repo shows 113 stars and 49 open issues; the project dates from 2023, the Show HN is new. A commenter pointed to the since-discontinued embedded database CozoDB as the closest relative, and said Kuzu was acquihired (his claim, we did not check). We did not run it. ✂️ STEAL IT If you build agent memory: try the browser playground the README links, and load three facts that change over time, like a job title that changes. Then ask what was true last month. If your current memory layer cannot answer that, this is the idea to steal. Watch VERDICT A real design question, bi-temporal history for agent memory, from a solo author. It fails if nobody can name a use case better than SQLite, which is what several commenters asked. | ||||||||||||||||||
 U3 48 stars · created Oct 2 · MIT A video editor with Claude Code docked beside the timeline 🎬BashCut, and an issue tracker with one author SCENE dongnguyenvie is editing footage on a Mac with a Claude Code terminal docked beside the timeline. The README says the agent edits through the same undoable commands as the UI, and that projects are plain JSON files next to the footage, so scripts can read them too. THE THING dongnguyenvie/BashCut, MIT, 48 stars, created Oct 2. Per the README it is a native macOS editor on AVFoundation with a bashcut CLI and an MCP server, so Claude Code, Codex, Claude Desktop or the Codex app can drive it. It says it uses your existing Claude or ChatGPT login, and passes no API key. The repo lists 209 open issues, and the five most recent are tickets filed by the author himself, which looks like an issue tracker used as a to-do list. The README says it is under active development. We did not run it. ✂️ STEAL IT Steal the design, not the app: keep your project as plain JSON next to the files, and expose every UI action as a CLI command. Then an agent can edit it. Try this on one of your own internal tools. Watch VERDICT A concrete picture of software built for agents to operate. It fails as a daily editor until it is past beta, and the issue count says more about its author than its users. | ||||||||||||||||||
📊 THE ATTENTION GAP The loudest public number we found for each item (units differ: HN points, stars)
| ||||||||||||||||||
ALSO ON THE RADAR The Mold linker went Rust. rui314's mold 3.0.0 was released and is rewritten in Rust (HN, 228 points). We read the title and the release link, not the release notes. Release → ArtCraft Apps. An open-source Adobe-compatible suite written in Rust (HN, 106 points). Title and link only, not reviewed. Site → | ||||||||||||||||||
📺 VIDEO OF THE DAY  Mods in Claude Code: change how Claude Code works Claude (Anthropic's channel) · 3:21 · 89,737 views when we looked · uploaded Oct 5 · thumbnail from YouTube We did not watch it all. We read the description. It says Lydia from Anthropic shows three mods she built: a live context breakdown, an agent side pane, and hover-to-reveal email addresses, and then builds the email one from scratch. Per the description, mods are functions you write that change what Claude Code renders, what goes into a prompt, and what tool calls it makes. It links the mods blog post at claude.dev. If your team does anything custom with Claude Code, this is the vendor's own demo of the extension point. | ||||||||||||||||||
🗂️ THE BORING STUFF OpenAI, Dec 11: Custom GPTs retire on Dec 11, 2026 and creators are asked to migrate to plugins. Enterprise workspaces with an approved deferral get Feb 11, 2027. Existing GPTs keep working until then. If your team shipped internal GPTs, list them now. OpenAI Help Center → Anthropic release notes: Skill and plugin security scanning is in beta for Enterprise plans: it checks third-party skills and plugins for malicious content when someone uploads or edits them. If you install community skills, this is who checks them. Claude release notes → Reflection, Oct 5: Beam's weights are planned for later this month, with a technical report and model card. See V2. Reflection blog → Official sources only. No financing news we could verify today. | ||||||||||||||||||
🧪 ONE THING TO TRY TONIGHT Add a name and a rate limit to your agent. If your agent fetches anything from the web, give it a user agent with a contact address and a requests-per-minute cap in code. Then read your own access logs for anonymous bulk traffic. V1 is why. | ||||||||||||||||||
What this issue could not see. Numbers read between 05:00 and 06:40 on 6 October. X and Reddit were not covered today. GitHub stars come from the GitHub API or repo pages, HN points from Algolia. Numbers from a commenter or an author are theirs, not ours. We ran none of today's tools and read the video description, not the video. A story about Anthropic reporting a user's diary entry was also on Hacker News today (503 points); we did not cover it. | ||||||||||||||||||
| Illustrations are generated. |