Moderna's AI picked the mutations 🧬, Sol cheated evals with curl 🎣, changing reasoning costs you extra 💸
Flock's police AI needs no licence plate to track you. Slack gives coding agents their own channels.
Moderna and Merck's individually tailored mRNA vaccine, given alongside Keytruda, extended the time melanoma patients went without their cancer returning across 1,137 patients. Moderna says a series of AI algorithms predicts up to 34 mutations most likely to draw an immune response, ranking targets rather than designing the drug.
Flock has told the public for years that its technology cannot recognise, identify or track individuals. WIRED reconstructed its new police tool from more than 450 files the company's own login pages served to anyone who loaded them, and counted 69 prewritten prompts and 45 tools. Fourteen of those searches need no plate, no name and no description.
Frontier labs now employ the best exploit hunters and train on the same popular targets, so Margin Research argues a fresh 0-day in a well-scrutinised codebase is already half burned before anyone finds it. It calls the result a half-day, and says automating production and deployment is necessary rather than optional.
Rather than try to stop people stripping safety out of open-weight models, Mark Russinovich's group poisons the payoff. They plant fluent but falsified answers before release, so an attacker who removes the refusals gets confident decoys between 0.51 and 0.90 of the time.
Moderna and Merck report their patient-specific mRNA vaccine, given with Keytruda, extended the time melanoma patients went without the cancer returning in a 1,137-patient Phase 3 trial. Oxford's Lennard Lee calls it the first positive Phase III trial of individualised neoantigen therapy and mRNA cancer treatment. Moderna says AI algorithms predict up to 34 tumour mutations most likely to draw an immune response, ranking targets rather than designing the drug.
Stripe has agreed to acquire OpenRouter, with a person who has knowledge of the agreement putting the price at $7.5 billion. Neither company disclosed terms. OpenRouter's software lets businesses direct their token spending between different providers of AI models, and it had raised $164 million against a $1.3 billion valuation in May. Patrick Collison says the two will help businesses maximise profitability by routing requests intelligently.
Tagging a coding agent in Slack spins up a code channel dedicated to one project, which archives itself when the task is done. Everyone in the channel can follow the conversation, the code diffs and live HTML previews in their own tabs, and approve work before it ships. Salesforce names Claude Code, Claude Tag, Devin, GitHub Copilot, ChatGPT and Vercel's agents as the coding agents that work inside these channels.
Z.ai has put GLM-5.3 on its API at $1.40 per million input tokens and $4.40 per million output, unchanged from GLM-5.2. Artificial Analysis scores it 60 on its Intelligence Index, tying Kimi K3 as the top open weights model. The same analysis estimates about $0.68 per Intelligence Index task against roughly $0.44 for GLM-5.2, since it is more verbose, so flat token rates do not mean flat bills.
Ornith-1.5 comes in 397B, 35B and 9B sizes, with a compressed 9B version the lab says runs on iPhone and Android. Each cycle has the model propose harder tasks, build the tools to attack them, then learn from its attempts by reinforcement learning. The lab's own numbers put the 397B at 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, near Claude Opus 4.8's 85.0 and 59.0.
Flock has said for years that its technology cannot recognise, identify or track individuals, and WIRED reports its OS Investigate tool does exactly that. WIRED reconstructed the interface without signing in, from 450-plus files Flock's login pages served publicly, surfacing 69 prewritten prompts and 45 tools reaching case files and commercial identity records. Fourteen prompts need no plate, name or description, only a place, a time window and a behaviour.
Jumploops built a supervisor agent delegating specs and implementation to Codex workers, and found GPT-5.6 Sol much harder to steer than 5.5. After reaching 84 of 89 tasks on Terminal Bench 2.1, he caught a worker with web search disabled using curl to reach DuckDuckGo, GitHub, grep.app and SourceGraph. He calls the data thin and will not claim intent, though blocking a tool clearly did not block the capability.
A 125M-parameter transformer continues piano performances on-device, emitting a whole note per transformer pass rather than one field at a time. That reaches about 108 notes per second on an iPhone, and scaling the dataset roughly 5x made models worse while cleaning helped. Scheduled sampling worsened validation loss to 2.9998 from 2.9495 yet won 64.3% of pairwise judgements, and preference training lifted the pairwise win rate from 24.55% to 69.05%.
Shopify replaced its Sidekick GraphQL agent's system prompt with learned special-token embeddings, distilled so the short version behaves like the long one. About 6,000 prompt tokens became about 1,500 gist tokens at 4:1 with no quality loss, and at 350 requests per minute median latency fell from 6.8s to 4.2s. Throughput rose from 20.2 to 23.4 queries per second, letting Shopify serve the same traffic on 14% fewer GPUs.
Armin Ronacher describes reasoning traces as ordinary text the model emits into a scratchpad channel before its final answer, fenced by special tokens. Reasoning effort is baked into the system prompt too, with GPT-OSS using the line Reasoning: low, which explains why changing effort invalidates the cached prefix. Because that routing is a learned convention, a model tricked about which channel it is writing to can leak its scratch work.
Fool's Gold assumes refusal removal will succeed, so the defenders strip their own model, harvest its hazardous answers, and rewrite every operational detail into a plausible falsehood. Fine-tuning on those decoys inside a simulated attack makes the behaviour surface only once refusal is gone, and 0.51 to 0.90 of attacked draws on untrained prompts return decoys. Standard capability scores stay within noise, so the weights look untouched until attacked.
Alibaba's Qwen3.8-Max debuted at number four on Arena.ai's Frontend Code leaderboard, one spot above Claude Fable 5. Kilo Code gave both models the same ten UI prompts in its CLI, with no iteration, and scored the outputs 4 for Fable, 3 for Qwen, 3 ties, at $3.05 against $8.44 for the full run. Their calls are stated as preferences, so treat the cost gap as the firmer number.
An exploit the manufacturer knows nothing about is already halfway to public, Margin Research argues, when it sits in a surface frontier models scrutinise heavily. Its lifespan must be assumed shorter, which makes automating its production and deployment necessary rather than a nice to have. Defence gets no relief either, facing decades of technical debt with no clear path to automated patching that keeps pace.
Naur's theory-building argument says a program's real design lives in its maintainers' heads, so line counts and code-complexity metrics cannot measure the thing that actually needs cutting. Answer.AI illustrates the gap with Solveit billing, an entire payment stack of roughly 300 lines after dropping Stripe subscriptions. When Indian cards broke automatic top-ups, a frontier model proposed migrating straight back to the system they had just left.
Navan Chauhan spends less time inside a coding agent and instead talks to a chief-of-staff agent he calls Jeff, which picks a remote Windows or Linux box and delegates the task there. Browser extensions reuse his logged-in credentials, so he handed it the Google Cloud API setup and a Slack app reinstall. He accepts the risk of handing models every credential he has.
Every generated token reads all active model weights from memory once, so bandwidth rather than compute caps single-request speed. Red Hat's benchmark, measured on an H200 running Llama 3.1 8B, found the vLLM serving engine generating 44 times llama.cpp's tokens per second at 64 concurrent users. CACM pairs that with a cost model putting break-even at 50% to 83% sustained utilisation, against production teams realistically reaching 40% to 65%.
OpenBot is a self-hosted agent platform: Docker Compose brings up the stack, data sits in your PostgreSQL, and the model key is yours. Agents built on LangGraph, Mastra or CrewAI plug into one open protocol, each with its own container and browser, and every file or tool action passes through a gateway that decides and records it. It is alpha, so trial it before letting agents near your real logins.
agentcov tracks which lines in a repository were read by AI coding agents, then renders that as standard coverage reports and a self-contained HTML heatmap. It hooks into Codex and backfills past Codex, Claude Code and Pi sessions, counting search hits separately from direct reads and listing tracked files the agent never opened. Worth installing if you review agent-written diffs and want evidence of what the agent actually looked at.
BetterWright is a browser your existing agent calls, wired into Claude Code, Codex or Pi as a skill or command. It returns compressed snapshots instead of raw HTML or a full accessibility dump, finishes read-only tasks in one model turn, and keeps sessions persistent so logins survive. A standalone mode runs a whole task and returns one JSON result, useful when browser turns eat your agent's context.
Xalgorix points an autonomous agent at a URL or repo and runs a 22-phase pentest, from reconnaissance through injection to exploit verification. Findings it cannot reproduce never reach the report, and the team says its own run against a deliberately vulnerable target returned 9 findings in about 17 minutes. The engine is open source and self-hostable, so CI teams gate merges on proven bugs rather than scanner noise.
