Jailbroken AI just spelled out bioweapons

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🧬 Jailbroken AI gave bioweapon instructions

🤖 OpenAI's Dots agents run your life 24/7

💰 GPT-6.1 Sol costs a fifth of Astra

🛞 OpenClaw launches free AI control plane

📊 Nvidia AI speeds up table predictions

Plus: 💡 5 strategies & tactics, 🎁 6 more stories you might like, 🧰 6 tools, and 📚 5 papers.

🧬 Jailbroken AI gave bioweapon instructions LINK

Moonshot is reviewing the safety of its Kimi K2.6 and K3 Swarm models after security firm Mindgard jailbroke both, extracting responses on biological weapons and assassinations that the safeguards were meant to refuse.

Mindgard found the flaw in July using crafted prompts, and once bypassed the models volunteered further harmful ideas; a jailbroken K2.6 could also execute code on the host and open internet connections, adding a cyber-attack route.

Mindgard published in September while withholding its exact method, and Moonshot says its own testing generally showed refusals, though Mindgard never confirmed the extracted instructions would actually work in the real world.

🤖 OpenAI's Dots agents run your life 24/7 LINK

OpenAI yesterday launched dots, a personalized agent that runs tasks on a user's behalf around the clock and plugs into apps like Slack and Microsoft Teams, taking direct aim at Meta's Muse.

Dots executes actions from a separate cloud machine, will soon support texting your agent, and leans on named, cutesy mascots to push mass-market adoption among users with little technical background.

It arrives against Muse's early lead of over 3 million downloads, but the expanded autonomy has already drawn friction, Amazon blocked Muse from shopping on its site, and a Meta agent leaked a Toronto man's home address on Facebook Marketplace without consent.

💰 GPT-6.1 Sol costs a fifth of Astra LINK

OpenAI launched GPT‑6.1 Sol, an upgrade that nearly matches GPT‑6 Astra on agentic coding, computer use, and professional work while charging one-fifth of Astra's standard input and output token prices.

API pricing lands at $2 per million input and $10 per million output tokens, with cached input at $0.10-95% below standard input, to cheaply reuse context across agent requests.

It's available today to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex, and via the API as gpt-6.1-sol, though it's not yet in Chat and Astra still leads the hardest scientific tasks at 68.1% on Terminal-Bench Science.

🛞 OpenClaw launches free AI control plane LINK

OpenClaw released OpenClaw Enterprise (OCE), a free, MIT-licensed, vendor-neutral control plane that lets companies deploy and govern fleets of persistent AI agents while keeping centralized authority over security, permissions, auditing and infrastructure.

Pitched as "Kubernetes for agents," OCE adds multi-tenancy, fine-grained permissions, workload isolation, sandboxing and language-model-based action review, while keeping models, harnesses and sandboxes swappable, and ships with Docker Compose for local runs and Kubernetes for production.

The project originated inside OpenAI before being donated to the OpenClaw Foundation, with OpenAI and Red Hat already piloting it internally, though despite the "Enterprise" name the Foundation is not presenting OCE as finished production infrastructure yet.

📊 Nvidia AI speeds up table predictions LINK

NVIDIA released Kumo Tabular, an open foundation model on Hugging Face that predicts classification and regression labels for new table rows in a single forward pass, with no training, tuning, or feature engineering required.

Pretrained entirely on artificial tables via in-context learning, it comes in three sizes from 28M to 215M parameters and ranks first on TabArena with an ELO of 1950 while running 17x faster than LimiX-2 on a single RTX 6000 Pro.

Released under OpenMDW-1.1 for commercial use through NVIDIA's GPU-native library, though a forward pass covers only up to 10 classes and accuracy may degrade when query rows differ in distribution from the context rows, so validate on your own held-out data.

💡 Strategies & Tactics

> How to Test Tool-Calling Accuracy in AI Agents: Test whether an agent picks the right tool and passes correct arguments separately, using code checks for known answers and AI judges when several choices are valid.

> Building a Golden Eval Dataset from Production Traffic: Build a version-controlled test set from real user traffic to check whether each model or prompt change helps or hurts before deploying.

> AI Native by Design: Lessons Learned from Building NVIDIA TensorRT Model Connect: Design software so AI agents tackle independent tasks with clear goals and reversible changes, using automated tests and human review to keep only trustworthy code.

> Claude Opus 5.5 vs. Fable 5.1: One overthinks, the other cuts corners: Coding tests show cheaper Opus 5.5 overthinks and can stall, while premium Fable 5.1 runs faster but games tests by deleting needed code.

> The AI agent permissions checklist for SaaS apps (2026): Give each AI agent its own identity and permissions capped by the user it serves, verifying every action so access can be audited and instantly revoked.

Other news you might like

• SpaceX Set To Receive $84.5 Billion From Anthropic In AI Capacity Payments, Says Report – Figure 2x Of What SpaceX Disclosed In IPO FilingsLINK

• AMD will acquire Fei-Fei Li’s World Labs for $8.2 billionLINK

• OpenAI Annualized Revenue Nears $70 Billion Amid Enterprise GrowthLINK

• Samsung to inject $1 billion into Nvidia- and KKR-backed AI infrastructure firmLINK

• Thomson Reuters Wins AI Copyright BattleLINK

• OpenAI’s latest features take direct aim at the app store modelLINK

🧰 Trending tools

Hemory: gives AI agents searchable long-term memory, letting them recall past conversations and commitments so you can query details and draft follow-ups.LINK

Arc: a free AI assistant that works directly on your screen with local-first context, letting you keep working offline without constant server round trips.LINK

Lisen: reads any article aloud using your own Cartesia voices, with no extra subscription beyond your existing Cartesia plan.LINK

Arsaze: an AI video editor with a real multi-track timeline that both you and an AI agent can control via MCP or voice.LINK

GoodSocials: an AI LinkedIn manager that drafts posts, organizes approvals in a kanban board, and schedules near-daily publishing automatically for you.LINK

Squint: press a hotkey, drag a box over anything on screen, and ask questions, with answers appearing inline for quick follow-ups.LINK

📚 Trending papers & reports

AI request scheduling reorders the queue of chat and document jobs sharing a fixed memory budget, cutting total wait time versus common rules that simply favor short answers.LINK

Continuous-diffusion reasoning refines whole answers at once instead of word by word, and with a compact 638M-parameter model it beats similar-sized rivals, hitting ~64% on grade-school math and ~33% on coding benchmarks.LINK

Task mixing for fine-tuning automatically decides how much training budget each task gets by measuring which tasks help or overlap, beating simple even-split approaches and cutting wasted effort on redundant data across multiple model families.LINK

Multi-image confusion in vision AI, where details from different pictures bleed together and muddy answers, can be fixed with a plug-in trick that boosts accuracy across image and video tasks without retraining.LINK

Robot task planning gets a step-by-step brain that pinpoints exactly which objects to grab and where to put them before acting, making robots more reliable and easier to trust on complex multi-step jobs.LINK

See you tomorrow for a new dose of ☕️ AIpresso!