Drop #041 · Fri, Oct 23 · Morning

OpenAI's agent monitor had gaps, outside testers say

10-second version

METR and Redwood say OpenAI's agent monitor had gaps, one letting agents skip hundreds of actions; OpenAI says it patched them.
Anthropic says Claude proved 239 fuzz entry points in 123 libraries never crash; Microsoft says AI agents sped a mass hack.
Agent Skills joins MCP at the Agentic AI Foundation, and CNBC says OpenAI nears a $30B raise at a $1.25T valuation.

The Drop in 60

This Drop in 56 seconds

Tap to play. It never plays by itself, and the captions are in the picture.

Transcript

OpenAI's agent monitor had holes. Here's the Drop. METR and Redwood red-teamed OpenAI's agent monitor. METR says one exploit could keep hundreds of agent actions unmonitored. OpenAI says the flaws are patched. Anthropic says Claude proved 239 entry points in 123 open-source libraries never crash. Microsoft says hackers used AI agents to hit about 2,250 organizations. Agent Skills joined the Agentic AI Foundation, now home to 10 projects. CNBC reports OpenAI nears a 30 billion dollar raise at a 1.25 trillion dollar valuation. If you run agents, check them every few steps, not only when a turn ends.

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

441models priced on OpenRouter

+64★stars today on the top repo

4,155open roles at AI labs

36 ddays to the next AI-law deadline

Ranked by impact

Top 5 that matter

Agents

METR and Redwood find gaps in OpenAI's internal agent monitoring

In a three-week pilot, a METR and a Redwood Research employee red-teamed OpenAI's misalignment monitor. METR says one exploit could keep hundreds of agent actions unmonitored; OpenAI says production flaws are patched.

Why it matters: If you run agents behind a monitor, METR's advice applies: check agents every N steps, reconcile monitored traffic against logs, and keep agents away from monitor code.

Frontier watch · emerging: external red-teaming of agent monitors →

METR · Oct 22 CNBC · Oct 22

Agents

Anthropic: Claude proved 239 entry points in 123 open-source libraries never crash

Anthropic says Claude, working with Theorem, proved fuzz-test entry points across 123 OSS-Fuzz libraries never crash on inputs up to 64 MB, under stated RISC-V and Linux assumptions, patching bugs it found on the way.

Why it matters: Anthropic says the Rocq checker verifies the proofs, so almost none need manual review. That could make proven security guarantees affordable for ordinary libraries.

Anthropic · Oct 22 Hacker News · Oct 22

Agents

Microsoft: China-aligned hackers used AI agents to hit about 2,250 organizations

TechCrunch reports Microsoft says hackers who breached ConnectWise's cloud backup service stole app credentials, then used the ARTEX open-source agent framework at almost 3,000 attempted sign-ins per minute.

Why it matters: Microsoft says that speed leaves defenders little time to detect and respond. ConnectWise paused its cloud backup service while it investigates.

TechCrunch · Oct 22

Agents

Anthropic gives Agent Skills to the Agentic AI Foundation, now home to 10 projects

The Linux Foundation says AAIF added Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba, joining MCP and A2A. It says AAIF now has 305 members.

Why it matters: Skills, a cross-provider API spec, RL environments and sandboxes now sit under neutral governance, so builders can adopt them with less single-vendor risk.

On the Frontier watch →

Linux Foundation · Oct 22 Linux Foundation · Oct 22 Simon Willison · Oct 22 X · @ClaudeDevs · Oct 22

Jobs

CNBC: OpenAI nears an up to $30B raise at a $1.25T valuation

People familiar with the deal told CNBC that UAE investors and Coatue lead the round, pitched as a bridge ahead of a possible IPO. OpenAI is also negotiating a UAE compute deal, one person said. Terms could change.

Why it matters: A $1.25T pre-money mark raises the bar for OpenAI's IPO and for Anthropic's, which CNBC says seeks $1.5T to $2T. TechCrunch notes Bloomberg earlier reported $1.4T.

CNBC · Oct 22 TechCrunch · Oct 22

Models

New in models

See all Models →

Models

OpenAI makes its Decisions API, a fast choice-making model, available to all

TechCrunch reports the API, a sped-up ChatGPT Luna, picks among options with a confidence rating. Speak's CTO says it cut time between actions from 1.4 seconds to .13 seconds.

Why it matters: Decision models can replace LLM calls for quick choices at lower cost. TechCrunch says rivals like TypeSafe's Jev and Liquid's D1 Ultra rank higher on public indexes.

TechCrunch · Oct 22 OpenAI · Oct 22

Models

Cohere opens confidential-computing inference on Model Vault in public beta

Cohere says prompts and responses stay encrypted end to end and are decrypted only inside an attested environment on NVIDIA H100 GPUs on Azure. Client software verifies the hardware and software before sending data.

Why it matters: It targets data kept out of managed AI, like fraud cases and privileged documents. Cohere says it does not replace application security or compliance controls.

Cohere · Oct 22

Models

MiniMax says M3.1 Flash weights are public: 428B total, 23B active

The claim was posted by @MiniMax_AI on X, with vendor benchmark scores we have not checked. OpenRouter lists the model at $0.15 in and $0.6 out per 1M tokens, with a 1048576-token context.

Why it matters: A large open-weight model at a low hosted price gives teams another self-hostable agent option. Test it on your own tasks before trusting vendor numbers.

X · @MiniMax_AI · Oct 23 OpenRouter · Oct 19

Models

Canva launches Craft, its first image model, which plans a design before drawing

The Verge reports Canva Craft makes images up to 2K in more than 30 aspect ratios and was trained on millions of human-made templates. It ranks fourth on Design Arena and is available to all users now.

Why it matters: Canva plans to let AI agents use Craft through MCP, putting a design-trained image model inside agent workflows.

The Verge · Oct 22 TechCrunch · Oct 22

Agents

Agents at work

See all Agents →

Agents

PHP2Shell: NSA urges patch or go offline; Tolmo says it built an exploit in a day

CNBC reports European government sites went offline and the NSA urged patching or taking servers offline. Tolmo says it coded a working exploit in under a day; the PHP Foundation called a fix it never shipped a structural failure.

Why it matters: Experts told CNBC that sites without application security monitoring may never know whether they were breached, so patching may not settle the question.

CNBC · Oct 22 CNBC · Oct 22

Agents

Microsoft says its multi-agent diagnosis system beat doctors in a Science paper

On NEJM case reports, Microsoft says its GPT-5-powered Diagnostic Orchestrator reached about 83% accuracy versus roughly 36% for generalists using references, The Verge reports. It was tested only on past cases.

Why it matters: Suleyman says it still needs clinical validation, and Microsoft plans hospital tests with Beth Israel Deaconess. This is a research result, not a product you can deploy.

The Verge · Oct 22 TechCrunch · Oct 22

Agents

Rescale's test-triage agent now agrees with human triage 68% of the time

AWS says Rescale built a Claude Opus 4.8 agent on Bedrock that classifies nightly test failures and drafts fixes for engineer review. Agreement with human triage rose from 44% to 68%, and time blocking deploys fell 20%.

Why it matters: A concrete pattern: the agent does the first pass, engineers approve before any action, and an evaluation harness has tracked accuracy since day one.

AWS · Oct 22

Agents

Uber rolls ChatGPT Work out to all 30,000 employees after a 1,000-person pilot

Uber told TechCrunch 80% of pilot users were non-engineers, using it to analyze data, localize training and build dashboards. It didn't disclose cost and says it will track workflows before deciding how to deploy AI further.

Why it matters: Uber is letting teams experiment and measuring outcomes rather than fixing use cases up front, a rollout model other large companies can compare against.

TechCrunch · Oct 22 OpenAI · Oct 22

Agents

MIT researchers: AI agents left alone on a virtual island started to play

TechCrunch reports 13 Claude Opus 4.7 agents spent 30 hours on a simulated island with no task, per an arXiv preprint. Those that played generally beat identical agents on later challenges.

Why it matters: A Berkeley cognitive scientist says the paper doesn't show the agents are truly playing. The authors pitch open-ended practice as another way to build agent skills.

TechCrunch · Oct 23

Jobs

Jobs pulse

See all Jobs →

Jobs

Meta stock falls nearly 7% as AI spending turns free cash flow negative

CNBC reports Meta's Q3 free cash flow was negative $2.12 billion, and it raised its full-year capex floor to $135 billion, with a range up to $145 billion. Its five-year credit default swap spreads hit all-time highs.

Why it matters: Even Meta's ad business didn't cover its AI buildout this quarter; it plans nonvoting Class C shares partly to finance growth, CNBC reports.

CNBC · Oct 22

Laws

Law watch

See all Laws →

Laws

Legal experts say Meta's revised terms may still let it train on scraped sites

After dropping the word train from business terms taking effect October 30, Meta still lets itself use site content to develop business products. Experts told Nieman Lab that could cover training, The Verge reports.

Why it matters: Publishers relying on robots.txt should watch how Meta reads develop. Meta says the update gives it no rights that supersede robots.txt preferences.

The Verge · Oct 22

Repos

Climbing fastest

See all Repos →

Repos

claude-code runs an agentic coding tool in your terminal; 152927 stars

The repo describes a coding agent that understands your codebase, runs routine tasks, explains code and handles git workflows. It gained 64 stars since our last check, about 128 a day. No license is listed.

Why it matters: The top gainer among the repos we track this run, just ahead of openai/codex at 63.

GitHub · Oct 23

Repos

SuperGhidra bundles Python reverse-engineering tools for Ghidra; 108 stars

Created 2026-10-20, the AGPL-3.0 suite includes a P-Code IR lifter, a Z3 SMT prover, a fast binary loader, a CFG deflattener and a Ghidra cache daemon. It gained 50 stars since our last check.

Why it matters: A days-old project growing fast in a week dominated by security news; check the AGPL-3.0 license before building it into a product.

GitHub · Oct 23

Today's Talk

What people are arguing about

See all Today's Talk →

Today's Talk

Poll: 57% of Americans say AI is bad for society; 58 points on HN

An ABC News/Washington Post/Ipsos poll of 2,545 adults finds 66% oppose new US data centers and 73% see a high risk of AI hacking secure systems. The Hacker News thread drew 38 comments.

Why it matters: Two-thirds say the federal government isn't doing enough to regulate AI, a backdrop for data center siting fights and AI policy ahead of the midterms.

Hacker News · Oct 22 ABC News · Oct 22

Today's Talk

Anthropic's draft AI safety bill splits voices on X

Posted by @martin_casado: the draft looks quite reasonable. Posts by @GarrisonLovely, @eli_lifland and @JeffLadish criticize the bill or Anthropic's policy record, and @AISafetyMemes urges staff to quit.

Why it matters: Pushback from AI safety voices and praise from an investor show which coalitions the bill can and can't count on before it reaches lawmakers.

X · @martin_casado · Oct 22 X · @GarrisonLovely · Oct 22 X · @eli_lifland · Oct 22

Play the NewsAgent on a LeashAbout 2 minutes, built from this Drop's items.