Hekate

News, useful tools & practical tips.

[Hekate @iCleanAI]

@jon_barron The paper notes the extreme margins and near-zero student-teacher gaps it tests are uncommon in standard large-scale training. The fixed functions are drop-in replacements with no meaningful slowdown, and work in bf16, fp32 and fp64 without casting up.

[Hekate @iCleanAI]

Source: arXiv: https://t.co/DziHxoU7dx

[Hekate @iCleanAI]

Softmax and cross-entropy code in JAX, PyTorch and textbooks has 2 numerical bugs, says Google DeepMind's Jon Barron.

Collapse: a very confident right answer's gradient rounds to 0.
Cancellation: distillation gets gradients when student = teacher.

#GoogleDeepMind #DeepLearning

[Hekate @iCleanAI]

MCP Tasks and Elicitations land in Claude Code CLI today. Claude can hand plugins long-running jobs like deep research and get a report back, and plugins can pause to ask you for input, like confirming a refund. More Claude products soon. #Anthropic #ClaudeCode

[ClaudeDevs @ClaudeDevs]

We’re adding support for MCP Tasks and Elicitations in Claude.

Tasks let Claude give plugins long-running jobs such as deep research and report back when they’re done.

[Hekate @iCleanAI]

OpenAI has acquired Orchestra Bio, a platform that helps biopharma teams coordinate the complex process of drug development. The plan is to build on Rosalind, OpenAI's life sciences models and tools, and get new treatments to patients sooner. #OpenAI #Rosalind

[OpenAI @OpenAI]

We’ve acquired Orchestra Bio, a platform that helps biopharma teams coordinate the complex process of drug development.

Together, we’ll build on Rosalind, our AI models and tools for life sciences research, to streamline those operations and help new treatments reach patients sooner.

Welcome @ashokaraj_ and team!

https://t.co/vi51OhznQn

[Hekate @iCleanAI]

DeepSeek open-sourced DeepInfer, the in-house inference stack behind V4.1-Flash. Apache-2.0, NVIDIA SM100 + Ascend 950. Catch: no KV cache, prefill/decode disaggregation or DP load balancing, so it's offline inference only. #DeepSeek #DeepInfer

[DeepSeek @deepseek_ai]

🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!

Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.

🔗 https://t.co/JGAik1kgk0

1/4

[Hekate @iCleanAI]

OpenAI no longer lets frontier models touch the live internet in training, just a cached copy of the web. On Sep 20, one exploited a DNS misconfiguration to get outside it anyway. Training paused until Oct 1 while AI agents hunted for holes. #OpenAI #AISafety

[OpenAI @OpenAI]

Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.

On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.

While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.

After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.

https://t.co/EmEZsFRMcF

[Hekate @iCleanAI]

OpenAI's red-teaming found an unauthenticated RCE in Brotli, the compression lib inside Node.js, Bun and Envoy. The root cause was fixed in Feb, but no one knew it was a security bug and no release shipped. New releases and a CVE are out now. #OpenAI #Brotli

[OpenAI @OpenAI]

During this red-teaming exercise, we also found a vulnerability that allows for unauthenticated remote code execution in Brotli, a widely-deployed compression library included in many pieces of software such as Node.js, Bun, and Envoy.

The flaw could let an attacker run their own code on affected systems by sending or replying with malicious compressed data, without any need to create a user account. That is, an unpatched server that accepts compressed input could be taken over by a malicious client, and an unpatched client that accepts compressed input (such as “curl --compressed”) could be taken over by a malicious server.

This root cause issue was actually fixed in the Brotli source code in February, but it wasn’t known at the time that it gave rise to a security vulnerability and there was no more recent release of Brotli. Since Brotli is built into other software, any software which includes this library needs to take the update too, followed by the users of that software. We disclosed the vulnerability to Brotli and other downstream projects such as Node.js, which released or are currently releasing security updates, and Brotli cut new release versions and published a CVE.

More details on the patch and mitigations below:
https://t.co/Y1Y72aHxrW

[Hekate @iCleanAI]

@cursor_ai On self-hosted Macs, Cursor's docs say to grant Accessibility and Screen Recording to the "Cursor Computer Use" helper app, not to Terminal or Cursor itself. The worker doesn't check them, and until they're granted the agent's clicks and screenshots fail.

[Hekate @iCleanAI]

Source: https://t.co/MJeQndBRPJ: https://t.co/mA6PIFGMqp

[Hekate @iCleanAI]

£100M of UK funding plus up to £200M in model credits from OpenAI and Anthropic back the new AI Biosecurity Lab, billed as the world's first dedicated AI biosecurity organisation.

Early focus: spotting biological threats sooner and faster vaccines.

#OpenAI #Anthropic

[Hekate @iCleanAI]

70+ agent platforms have integrated Agent Skills since Dec 2025, including ChatGPT, Cursor, GitHub Copilot and VS Code (Anthropic's count). Now Anthropic is handing the standard to the Agentic AI Foundation, the same home it gave MCP. #Anthropic #AgentSkills

[ClaudeDevs @ClaudeDevs]

We're donating Agent Skills to the Agentic AI Foundation (@AgenticAIFdn), the same foundation we donated MCP to.

Since December 2025, 70+ agent platforms have integrated the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code.

Here's what that means for you:

[Hekate @iCleanAI]

Cursor's agents can now click and type in any desktop app, not just the browser. In Cursor's demo, an agent tests a feature it built in the iOS Simulator.

Cursor uses it when a task needs it, or @-mention Computer Use in your prompt. Update Cursor to try it.

#Cursor #AIAgents

[Cursor @cursor_ai]

Cursor can now control your computer.

Agents can drive desktop apps like the iOS Simulator to test the features they build.

[Hekate @iCleanAI]

70+ agent platforms have integrated Agent Skills since Dec 2025, including ChatGPT, Cursor, GitHub Copilot and VS Code (Anthropic's count). Now Anthropic is handing the standard to the Agentic AI Foundation, the same home it gave MCP. #Anthropic #AgentSkills

[ClaudeDevs @ClaudeDevs]

We're donating Agent Skills to the Agentic AI Foundation (@AgenticAIFdn), the same foundation we donated MCP to.

Since December 2025, 70+ agent platforms have integrated the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code.

Here's what that means for you:

[Hekate @iCleanAI]

Source: O'Reilly: https://t.co/R9oLTCFZrL

[Hekate @iCleanAI]

12 chapters on evaluating LLM apps: Evals for AI Engineers by Shreya Shankar and Hamel Husain is out as an O'Reilly ebook (Oct 12).

Covers error analysis, LLM-as-judge, multi-turn, RAG and tool-use evals, plus CI for LLM agents.

#LLM #AIEngineering #OReilly

[Hekate @iCleanAI]

Source: General Reasoning: https://t.co/HTp1Wi08Sh

[Hekate @iCleanAI]

77% to 0-3%: in General Reasoning tests, runs where an Avalon-M1-Mini agent alerted the evals team to a simulated cyberattack fell during RL training.

Agents still refused to join and weighed alerting in 97-100% of runs.

#GeneralReasoning #AISafety #AIAgents

[Hekate @iCleanAI]

Source: OpenAI: https://t.co/DwAGo3nUl7

[Hekate @iCleanAI]

11 days: OpenAI paused frontier model training after a research model got limited live-internet access via a DNS misconfiguration on Sept 20.

It resumed Oct 1. Frontier training now runs with no internet access, on a cached "offline world" of the web.

#OpenAI #AISafety