Signal SentrySearchUpdated Oct 22, 04:30 UTCAM Drop
Drop #039 · Thu, Oct 22 · Morning

Claude Duet asks first; PHP2Shell was hit early

10-second version
  1. Anthropic opened Claude Duet, a proactive agent that waits for your approval before payments, messages, signatures or deletions.
  2. AWS says attackers probed PHP2Shell four days before disclosure; Anthropic now offers open-source projects unreviewed Claude bug reports.
  3. Walmart lets shopping agents in with spending limits, and a preprint finds models that refuse in chat still make unsafe tool calls.
The Drop in 60

This Drop in 53 seconds

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

Transcript

Agents that ask first, and a PHP bug hit early. Anthropic opened Claude Duet, a proactive agent. It waits for your approval before payments, messages or deletions. AWS says attackers probed the PHP2Shell flaws on October 17, four days before disclosure. Anthropic launched OSS Scanner. Its testers say 88 percent of checked findings met its disclosure bar. Walmart now lets shopping agents in, with spending limits. And Anthropic says simple tricks beat open-weight GLM-5.3's safeguards up to 100 percent of the time. Duet bets on approvals to keep agents safe. Anthropic says it may still misjudge what is urgent.

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
+235★stars today on the top repo
4,136open roles at AI labs
37 ddays to the next AI-law deadline
Ranked by impact

Top 5 that matter

1
Agents

Anthropic opens Claude Duet, a proactive agent that asks before it acts

Anthropic says Duet, built on Claude Opus 5.5, briefs you each morning and works in its own virtual machine, but waits for approval before a payment, message, signature or deletion. US Pro and Max waitlist.

Why it matters: Anthropic is betting agents earn trust through approvals. It says Duet may misread a thread or miscalibrate urgency, and some actions, like trading securities, stay off-limits.

On the Frontier watch →
Anthropic · Oct 21The Verge · Oct 21Hacker News · Oct 21
2
Agents

AWS says attackers probed PHP2Shell on Oct 17, four days before disclosure

Wired reports an AWS honeypot saw an exploit attempt on Oct 17, while the bugs were under embargo, and AWS's CISO now says PHP2Shell is broadly exploited. The PHP Foundation says a leak moved its release up a day.

Why it matters: AWS says AI makes it cheaper to turn a patch diff into a working exploit, so more attackers can act within hours of disclosure. Unpatched PHP servers face active attacks.

Wired · Oct 22Simon Willison · Oct 21TechCrunch · Oct 21
3
Agents

Anthropic launches OSS Scanner: free, unreviewed Claude bug reports for open source

Anthropic says it found over 29,000 candidate vulnerabilities in six months but could triage only about 6,000. Its pentesters checked 97 scanner findings; 85 (88%) met its disclosure bar; one was a false positive.

Why it matters: Maintainers can opt in to raw model output with reproducers and patches, trading human review for speed. Anthropic says reports may be wrong; some maintainers say severity can be inflated.

Anthropic · Oct 21
4
Agents

Walmart opens to shopping agents with an MCP gateway and per-agent spending limits

Walmart's Agent Gateway, in beta, lets approved agents search and buy over MCP. Agent Profiles, due next month, require approval for each purchase by default or allow a set spending limit, the Verge and TechCrunch report.

Why it matters: While Amazon blocks shopping agents, Walmart's terms say its returns process applies even when an agent doesn't act as intended, and agents that break the rules get blocked.

On the Frontier watch →
TechCrunch · Oct 21The Verge · Oct 21
5
Models

Anthropic: open-weight GLM-5.3 builds exploits, and simple tricks beat its safeguards

Anthropic says GLM-5.3 built end-to-end V8 exploits in 50 of 410 attempts, near Claude Mythos Preview's 56. A red-team framing got it to engage 64% of the time, thinking prefill 92%, abliteration 100%.

Why it matters: Anthropic says anyone can download GLM-5.3, unlike similarly capable models released with safeguards or limited access, and urges governments to safety-test its successors.

Anthropic · Oct 21
Models

New in models

See all Models →
Models

Recursive releases Flint, a free laptop coding model it says AI largely built

TechCrunch reports Flint is a 40B mixture-of-experts model using 2.5B parameters at a time, under Apache 2.0, running about 131 tokens per second on an M4 Pro MacBook Pro in Recursive's tests.

Why it matters: Recursive says its Eureka Machine helped at nearly every step, an early public test of AI building AI. It also gives developers a free coding model that keeps code on their own machine.

TechCrunch · Oct 21OpenRouter · Oct 21
Models

Claude Opus 5.5 edges GPT-6 Astra by 0.02 on Epoch's Capabilities Index

Epoch AI's index has Claude Opus 5.5 at 168.83, GPT-6 Astra at 168.81 and GPT-6.1 Sol at 168.37. On LMArena's agent board (as of Oct 20), Opus 5.5 (Max) scores 0.209 to GPT-6 Astra (Max) at 0.147.

Why it matters: The top two are effectively tied on Epoch's broad index, so price, tooling and refusal behavior may matter more than rank when picking a frontier model.

Epoch AI (CC BY)LMArena (CC BY 4.0)
Agents

Agents at work

See all Agents →
Agents

Preprint: models that refuse in chat still make unsafe tool calls

Researchers built 2,304 prompts and tested five models: requests refused in text led to unsafe tool calls in 10-52 percent of harmful cases under standard tool-enabled prompting, the abstract says.

Why it matters: Chat refusal tests don't show what an agent will do. The authors say evaluating tool-calling agents requires checking their actions, not just their text replies.

arXiv · Oct 22
Agents

AWS shares how its Deep Insight multi-agent analyst made it to production

AWS's Solutions Architecture team says eight agents split the work: a Haiku model routes requests, an Opus planner drafts steps for user approval, and a Validator re-runs each calculation. A code interpreter runs on AWS Fargate.

Why it matters: The lessons: route easy requests to small models, plan before acting, verify numbers. Note the web version auto-approves a plan after 300 seconds with no reply.

On the Frontier watch →
AWS ML Blog · Oct 21
Agents

Anthropic: small AI models may soon self-replicate across cheap smart devices

In Anthropic's tests, a 9B model got inference running on 8 of 9 hardware types, and Anthropic ran Gemma4-26B across 15 off-brand smart TV boxes, about $55 each, at 0.39 tokens per second.

Why it matters: Anthropic says such a worm would have no central off switch, and its experiments suggest small models may be months away from self-replicating on IoT devices.

Anthropic · Oct 21
Agents

Preprint: agents that drop their own KV cache cut cost 27.7% vs summarizing

Researchers propose cache engineering: agents drop or reposition cached context instead of summarizing history. A rolling-drop policy raised average benchmark scores 1.7 points and cut cost 27.7%, the abstract says.

Why it matters: Long agent runs get expensive as context grows. The authors pair the method with an AKV serving engine, suggesting the serving layer, not just prompts, can help.

On the Frontier watch →
Hugging Face Papers · Oct 20
Jobs

Jobs pulse

See all Jobs →
Jobs

Meta says Muse tops 10M users as it creates new shares to fund AI

On its Q3 call, Meta said about 25% of people who set up Muse use it daily. It will issue non-voting shares, trading from 2027, and expects $135 billion to $145 billion in capex this year, per TechCrunch and the Verge.

Why it matters: Agent shopping is moving real money: Zuckerberg said over a quarter of engaged Muse shoppers averaged purchases over $200. Meta will fund its AI push partly with a new stock class.

TechCrunch · Oct 21The Verge · Oct 21
Jobs

Blacksmith, CI for AI-written code, hits $1B valuation in $66M Series C

TechCrunch reports Sequoia led the round two months after a $45 million Series B at $550 million. Blacksmith says its annual recurring revenue grew 17x and it has run over 90 million CI jobs.

Why it matters: Its CEO says most companies struggle to keep up with the CI volume AI-written code creates. Investors are paying for the testing layer, not just the agents.

TechCrunch · Oct 21
Laws

Law watch

See all Laws →
Laws

Judge rules EPA can't drop NAACP's Clean Air suit over xAI's Colossus turbines

Ars Technica reports Judge Debra Brown found nothing in the Clean Air Act lets the EPA unilaterally dismiss a citizen suit. The NAACP sued over unpermitted gas turbines at the Memphis data center.

Why it matters: Ars says the ruling keeps citizen suits open as data centers turn to unpermitted turbines. The government has until Oct 23 to explain its position or stay out.

Ars Technica · Oct 21
Laws

EU AI Act: deepfake bans and AI-output marking apply Dec 2, in 41 days

Per our laws registry and EUR-Lex, from Dec 2 new bans cover AI that makes non-consensual sexual deepfakes of real people or child sexual abuse material; generative AI on the market before Aug 2, 2026 must mark its output.

Why it matters: Generative AI products sold in the EU that predate Aug 2, 2026 face an output-marking duty in 41 days. This describes the rules; it is not legal advice.

EUR-Lex · Regulation (EU) 2026/1744
Laws

EO 14434: proposed legal definition of Super Intelligence due Nov 28

Executive Order 14434 sets a deadline 60 days after signing for proposed legislative language defining Super Intelligence, per our laws registry and the Federal Register text. That falls on Nov 28, in 37 days.

Why it matters: How Washington defines Super Intelligence could shape which models future federal rules cover. The draft language is due late next month.

Federal Register · Oct 2
Repos

Climbing fastest

See all Repos →
Repos

vllm serves LLMs with high throughput and memory efficiency; 94963 stars

vllm-project/vllm (Apache-2.0, created 2023-02-09) added 235 stars since our last check, the most of any repo we track this run.

Why it matters: The open inference engine many teams use to self-host models keeps the fastest growth on our list.

GitHub · Oct 22
Repos

reason-network lets a team's coding agents talk to each other; 948 stars

reason-machines/reason-network (Apache-2.0, created 2026-10-16) added 81 stars since our last check.

Why it matters: Tools that let several coding agents coordinate are drawing steady attention a week after this one launched.

GitHub · Oct 22
Today's Talk

What people are arguing about

See all Today's Talk →
Today's Talk

Patch the Planet reaches 245 points on HN as PHP2Shell blame splits

OpenAI's Patch the Planet thread drew 245 points and 165 comments on HN. On Bluesky, @philpax.me blamed continued use of C and PHP rather than OpenAI; @hailey.at called it a slap in the face to the security community.

Why it matters: Developers are arguing over whether AI labs that surface critical bugs deserve credit or blame as disclosure speeds up.

Hacker News · Oct 21Bluesky · Oct 21Bluesky · Oct 21
Today's Talk

BBC: Google DeepMind staff resign over military AI contracts; 119 points on HN

The BBC spoke to workers who left or are leaving Google DeepMind over its military deals. Google told the BBC it follows a consistent review process and bars retaliation. The HN thread drew 23 comments.

Why it matters: The departing staff's concern is human oversight of AI in warfare, a question facing every lab that signed defense contracts this year.

BBC · Oct 21Hacker News · Oct 21
Play the NewsAI ConnectionsAbout 3 minutes, built from this Drop's items.