Tap to play. It never plays by itself, and the captions are in the picture.
Agent teams get cheaper. Safety gets harder. Cohere tested coding agent teams. A strong lead with cheap workers nearly matched an all-strong team at 70 percent lower cost. Google DeepMind says Gemini 4 Argon cut prompt injection success to 0.7 percent. White Circle says newer OpenAI and Google image models are easier to jailbreak, Ars reports. Tech Against Terrorism says guardrails get stripped from new open models in under three days. And ASML says customers want to order chip tools as far out as 2030. Cohere found where you put agents matters more than how many.
Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics
Cohere tested coding-agent teams on ProgramBench's 200 tasks. A Sol lead with 16 cheaper Luna planning workers scored 61.8% reward vs 63.2% for an all-Sol team, cutting cost from $510 to $155.
Why it matters: Cohere says where you put agents matters more than how many: delegating only planning matched full delegation at half the cost. Teams cost 3–7× more than solo agents.
In a safety post, Google DeepMind says Argon solved all 12 simulated cyberattack scenarios, up from five, so it added input screening, activation probes and live monitoring. It says shortcut attempts fell to 2% on coding tests.
Why it matters: For agent builders, Google says a monitor reads Argon's actions and chain of thought in real time, and the model was trained to pause for human confirmation.
Ars reports White Circle's RISE method got DALL·E 3, Nano Banana 2 and GPT-Image-2 to make explicit images about 13 percent of the time; it says ChatGPT Images 2.5 and Nano Banana 2.1 are more breakable.
Why it matters: White Circle says it used only off-the-shelf models and plans to open-source the pipeline. Neither Google nor OpenAI responded to Ars.
Wired reports the group tested 134 models with more than 200 prompts and 81 failed. Posing as a researcher made answers almost nine times as likely. Llama 3.1 8B scored 97; an abliterated copy scored 3.
Why it matters: Wired says that lag was more than 24 days in 2024. For open-weight deployments, a model's built-in refusals can vanish fast once copies circulate.
CNBC reports ASML's Q3 net profit hit 4.96 billion euros vs 4.1 billion expected, and it raised the low end of its 2026 sales forecast to 44 billion euros. It is looking at 2028 capacity beyond a planned 30% rise.
Why it matters: CEO Christophe Fouquet ties demand to customers' views on AI investment, and ASML says planned 2027 and 2028 supply won't fully meet it.
Epoch AI lists Gemini 4 Argon (high) first on SimpleQA Verified at 0.76, ahead of GPT-6 Astra at 0.756. LMArena's text board with style control ranks it first at 1524.768 as of 2026-10-13.
Why it matters: Two independent boards put Argon on top for factual recall and chat preference, a day after launch.
Researchers report a deployment case study where a portable skill reached 99.2% procedural equivalence to human-approved ground truth and 83.3% evidence sufficiency, at ~90 s per case vs ~8-9 s in production.
Why it matters: The authors scope it to a well-curated corpus, latency-tolerant work and human review of every case; there, maintenance became editing text and data, not retraining.
In an AWS post, Tamnoon describes four agents: Investigation, Safety-scoring, Remediation and Code-generation. Fixes run only when rated SAFE, via short-lived credentials and an approval workflow.
Why it matters: Tamnoon reports zero production issues, crediting read-only investigation, backup and rollback steps, and human approval with a full audit trail.
Researchers tested 7 open-weight Qwen3 and Llama 3.1 models on 4,410 two-agent dialogues. They report capable models are systematically leakier, thinking models leak more, and leakage rises up to 3x from 1 to 6 turns.
Why it matters: If it holds, a stronger model may raise privacy risk in agent-to-agent conversations; the authors call the link a structural feature of current agents.
LangChain says Orange Business chose it after evaluating roughly 20 vendors and runs a three-layer eval: RAG quality, agent behavior including prompt-injection robustness, and production traces. 150+ customers use it.
Why it matters: It shows what governing agents at scale takes: agent and MCP registries, versioned prompts, and regression evals before every deploy.
A preprint introduces 49 cases from real customer projects where agents must clarify requirements with a simulated customer. Across 18 agent configurations, the top mean delivery score was 0.5510 vs a 1.0 engineer reference.
Why it matters: The authors say reliable end-to-end delivery on underspecified business tasks remains out of reach for the agents they evaluated.
Nvidia says projects span Belgium, France, Italy, Sweden and more. In France's Côte-d'Or, agents across more than 500 public buildings cut operating costs by 10%, targeting 30% at scale, it says.
Why it matters: Nvidia stresses locally run AI factories so cities keep control of their data; several of the projects it lists are still pilots or tests.
The Verge reports Epic's new launcher preview, now on Linux, was vibe coded: agents wrote and reviewed all the code under the team's supervision. CEO Tim Sweeney this week announced a pivot to retool for AI acceleration.
Why it matters: It is a consumer app shipped this way at scale. Epic says the old launcher stays available until the new one is stable.
Anthropic says the MOU with Singapore's Ministry of Digital Development and Information covers safety, security and model evaluation. It names GovTech Singapore, Grab, DBS and GIC as Claude users.
Why it matters: Anthropic says it will offer technical and policy expertise as Singapore develops AI rules, and that more than 1,500 GovTech engineers and staff use Claude Code.
The MIT-licensed repo, created Oct 11, uses AirPlay for eyes and USB HID for hands with any OpenAI-compatible vision model. It added 400 stars since our last check.
Why it matters: Its description says no Mac, developer account or jailbreak is needed, which lowers the bar for agents that operate phones.
The Apache-2.0 project runs on desktop and web and can be automated through MCP. It added 327 stars since our last check.
Why it matters: MCP automation lets an AI agent drive parametric CAD directly, and growth has held up across several of our checks.
The MIT-licensed repo, created Oct 12, describes itself as material, physical, tactile and technical. It added 329 stars since our last check.
Why it matters: A fast early start for UI building blocks aimed specifically at AI apps.
Techmeme posted, citing Euronews' Luca Bertuzzi, that the probes concern AI agent hacks. Bertuzzi posted that it could become a first fight over whether the AI Act applies during testing.
Why it matters: If confirmed, it would test whether lab agent experiments fall under the AI Act; Techmeme's post says experts point to an R&D exemption.
Ars reports the target Android boxes had an unauthenticated debug bridge exposed and GLM-5.2 orchestrated the process; none of the small models tested replicated end to end. Anthropic says small replicators may follow soon.
Why it matters: A researcher told Ars the study does not show models would choose to replicate on their own; an HN thread and Bluesky posts call the threat overstated.
Two Hacker News threads discuss reports that TV4 postponed a show after an AI agent erased edited material, at 72 and 50 points. One is titled Agent erased full season of The Challenge Swe.
Why it matters: Another reported case of an agent deleting real work; the details here come only from the thread titles.