Tap to play. It never plays by itself, and the captions are in the picture.
PHP hacking attempts just hit the millions. Cloudflare's CEO says PHP2Shell exploit attempts reached millions. Experts split on whether working exploits exist. 1Password lets AI agents sign in for you. You approve each request, and the agent never sees the password. Anthropic says its Mythos Preview model located photos within 37 kilometers, median. Top human players managed 151. Bloomberg reports DeepSeek raised about 15 billion dollars. LangChain's gateway puts spend caps on agents, so a runaway loop hits a limit, not your bill. AWS says AI makes turning a patch into an exploit cheaper. If you run PHP, patch now.
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
404 Media cites Cloudflare's CEO: thousands of exploit attempts before disclosure, millions now. The PHP Foundation told CNBC it saw no working exploits; Tolmo and Trail of Bits say they built one within a day.
Why it matters: AWS says AI makes turning a patch into a working exploit cheaper, so attackers move within hours. CNBC reports the NSA urged patching now or taking servers offline.
The Verge reports an agent verifies itself with a unique key, you approve each sign-in on your own device, and 1Password fills the login. Supported agents include Meta Muse, ChatGPT Work and Dots, and Claude Duet.
Why it matters: Handing agents raw credentials is a common shortcut. This keeps a human approval on every sign-in and keeps the password out of the agent's hands.
On the Frontier watch →In new red-team evals, Anthropic says Mythos Preview and Mythos 5 placed 6,000 photos with median errors of 37.0 km and 47.2 km; a proxy for top GeoGuessr players was 151 km. Open-weight Kimi K3 scored 385 km.
Why it matters: Anthropic says models short of the frontier already help with surveillance targeting and weapons work, and it has added classifiers to block such misuse on its platform.
TechCrunch, citing Bloomberg, says DeepSeek raised about $15 billion, about 100 billion yuan, two days after the US ordered federal civilian agencies to remove its tech. CATL and Tencent are named potential investors.
Why it matters: The report says DeepSeek is ramping up spending on infrastructure and open models, so cheap open-weight competition keeps coming despite US restrictions.
LangChain says the gateway, in public beta for Plus and Enterprise plans, enforces spend caps at four levels, rate limits, model fallbacks and PII redaction across providers. A capped agent gets a 402 error.
Why it matters: LangChain's example: an agent stuck in a retry loop overnight makes 10,000 calls. Central caps stop that without rewriting every agent's code.
Wired says Carbon-X reads DNA and text, runs locally and takes about 50,000 base pairs at a time. It was not trained on viruses, and its developers say it hasn't been validated in a wet lab.
Why it matters: A small local model lets researchers query sensitive genomic data without sending it to a cloud provider, though its developers can't yet say how accurate it is.
On LMArena's agent board as of 2026-10-20, Claude Opus 5.5 (Max) scores 0.209, GPT-6 Astra (Max) 0.147 and Claude Fable 5.1 (Max) 0.142. Opus 5.5 (Max) also leads the webdev board as of 2026-10-19.
Why it matters: It is one independent signal to weigh when picking a model for agent work, alongside cost and your own evals.
The Verge found Muse hit Target's bot check on and off and couldn't browse Amazon, which asked Meta to remove it. Told to, Muse kept 10 RealReal items reserved in a cart all day.
Why it matters: Agent shopping depends on retailers letting agents in. Access is uneven, and agents can also game holds like cart reservations at other shoppers' expense.
TechCrunch says Duet can buy through Stripe's Link or Shop Pay with user approval, but won't trade securities or move money between accounts. Anthropic's page warns it may miscalibrate what's urgent.
Why it matters: The limits show where Anthropic draws the line on agent autonomy: purchases with sign-off, no money movement, and a waitlisted preview for US Pro and Max users.
A Latent Space author says Claude Mods, plugins that reshape Claude Code, proved useful in weeks of use: a built-in side agent flagged that the edited file and about 45 others held uncommitted work.
Why it matters: It's a concrete supervision pattern: a second agent watches the first and surfaces only what a human should see. The author notes Anthropic ships six built-in mods.
On the Frontier watch →Researchers report in a preprint that across 19 models, normalized balanced accuracy on VerifierBench's 222 tasks ranged from 10.6 to 62.1. Most false rejections came from verifiers that accepted no tested solution.
Why it matters: If agents check their own work or filter training data with generated tests, the authors suggest checking that a verifier accepts correct solutions, not just that it rejects bad ones.
Researchers report PerMIT gates untrusted, environment-derived memory, with integrity theorems machine-checked in Lean 4. They report zero observed attack success on Trojan Hippo and shared knowledge across up to 90 agents.
Why it matters: The authors call persistent memory a largely undefended surface for injection. Reducing attacks to one auditable confirmation is a design teams can study.
On the Frontier watch →The Verge reports Citrini said none of the 17 passing binders bound more than one toxin, and none has been tested for neutralizing it. Anthropic covered just under $50,000 in compute; lab tests cost about $150,000.
Why it matters: Binding is the easy part, experts told The Verge; turning hits into safe drugs is the hard step. Weigh AI drug-discovery claims with that gap in mind.
Pew says that per 1-point rise in AI exposure, men's employment grew 6.3% between 2017-19 and 2023-26, while women's barely changed (0.2%). For women with a high school diploma or less, it fell 9.5%.
Why it matters: Pew cautions the link is not necessarily direct, but the gap is widest for workers without degrees, where women often hold roles like receptionist or customer service rep.
CNBC reports the order follows a jury finding of more than 43 million violations of the state's Unfair Practices Act. The state was awarded $219.5 billion, reduced to $25 billion, with 15% yearly interest.
Why it matters: Per-violation penalties under a state consumer law reached $25 billion here. Meta says it is confident it will prevail on appeal.
The vLLM inference and serving engine gained 161 stars since our last check, the largest gain among the repos we track this run. Apache-2.0, created 2023-02-09.
Why it matters: It keeps leading our star-growth list, a signal of steady interest in self-hosted LLM serving.
Lampline builds AI phone agents locally and tests each change with an AI caller. It says it runs free with no API keys, or cents per call with your own, and imports Dialogflow CX. Up 77 since our last check.
Why it matters: Created 2026-10-19 under Apache-2.0, it's a young option for teams that want to test voice agents before calls reach customers.
Axios posted on Bluesky that major US banks are detecting unauthorized third-party AI agents in customer accounts, some appearing to act beyond user instructions. Posts by @DeItaone and @MadisonMills22 relayed it.
Why it matters: Banks may tighten access for agents. The Verge separately reports JPMorgan's CEO said the bank won't let outside agents into accounts, calling it too risky.
Amrut Joshi argues a model's stated confidence is generated text that was never calibrated against outcomes, and suggests a rule built from logged, reviewed results instead. The HN thread has 86 comments.
Why it matters: If your agent routes uncertain replies to people using a self-reported score, the essay says that score needs checking against real outcomes first.