PHP2Shell: AI-assisted research uncovers critical flaws at the heart of PHP
Researchers disclosed two memory-corruption vulnerabilities in PHP's core engine, named PHP2Shell (CVE-2026-109618 and CVE-2026-109619), affecting roughly 70% of the web's sites. OpenAI's "Patch the Planet" programme (using GPT-5.6 Sol) and security firm Wiz (using Gemini 4 Argon) independently found the flaws, which allow pre-authentication remote code execution against WordPress, Drupal and Joomla and Heartbleed-style memory disclosure on any PHP site. The NSA urged organisations to patch immediately or take systems offline, Cloudflare deployed stopgap firewall rules, and a fix shipped after roughly two months of coordinated work between the researchers, PHP maintainers and governments.
Anthropic opens a research preview of Claude Duet, a proactive AI agent
Claude Duet works from its own cloud computer and browser rather than a user's device, sending a morning brief across a person's connected apps and checking back every few hours. It shows its plan and waits for approval before sending a message or spending money, and keeps some actions, such as trading stocks or moving money between accounts, off-limits regardless; Anthropic says it can still misjudge urgency and remains weaker at personal errands such as booking tickets. Access is starting with US Pro and Max subscribers.
Sam Altman sketches three possible endgames for the AI safety race
In an internal talk reported by The Information, the OpenAI chief set out three ways the industry could end up managing self-improving AI: a global watchdog body modelled on the International Atomic Energy Agency, a single merged lab serving partner countries (the "Demis plan"), or one lab racing ahead to disable its rivals' computing, which Altman called "terrible". He reportedly wants OpenAI to commit to a path, including a "speed limit" on self-improvement, within a year.
Study of 331 Microsoft developers finds AI's speed gains outrunning their judgment
Researchers from Microsoft, Oregon State University and DX sorted what AI newly lets developers do into three zones: Faster, familiar work sped up (45% of responses); Further, work previously out of reach or not worth the effort (37%); and Afresh, new work built around directing and checking agents (12%). The study warns of a "hollow orchestrator" effect, in which a developer approves more AI output each month without their own judgment keeping pace with it, and flags a small rise in DX's Q3 change-failure-rate data as an early signal worth watching.
DORA's Nathen Harvey: AI amplifies whatever performance culture already exists
Drawing on the 2025 DORA research, Harvey argues AI adoption does not fix organisational dysfunction by itself: it scales whatever habits, internal platforms and guardrails are already in place, for better or worse.
Radical Numerics claims a world first with an FDA-cleared, AI-designed virus
Radical Numerics says the FDA cleared a virus its Omnii-1 AI designed for a patient with a drug-resistant infection, which it is calling the first AI-designed medicine to reach a patient.
OpenAI's GPT-Rosalind links mouse-brain gene data to research literature in minutes
Thousands of agents cross-referenced gene-expression results from millions of mouse-brain cells against the scientific literature in 32 minutes, proposing testable hypotheses such as a link between disrupted DCTN1 and impaired cholesterol transport. The output is hypotheses for researchers to test, not validated findings.
AI reads routine pathology slides to flag ovarian cancers that respond to targeted drugs
King's College London researchers trained an AI on tissue slides from patients with known genetic-test results to spot ovarian tumours with a broken DNA-repair mechanism, often linked to BRCA1 or BRCA2 changes, which makes them more responsive to certain chemotherapies. It caught about 91% of affected tumours, potentially offering a cheap first screen ahead of the expensive genetic test it normally takes to find out.
Ex-DeepMind researcher launches Metis Reasoning to make chatbots check their own answers
Thore Graepel, formerly of DeepMind, has started Metis Reasoning, applying techniques from AlphaGo-style self-play to get chatbots to verify their own reasoning against evidence.
Skyfall AI launches SKY-001 to speed up 3D design work
Skyfall AI's SKY-001 drafts 3D designs for parts and buildings, which the company says it does more than 51 times faster, and more cheaply, than frontier general-purpose models.
Zoom to add live deepfake detection to calls
Zoom has announced deepfake-risk detection that analyses live video and audio during a call for signs of alteration and shows a confidence indicator, due in December.
An old smartphone becomes the brain of a small walking robot
UCLA researchers built PhoneBot, a roughly 19-inch, 4lb walking robot that uses a regular smartphone's camera for navigation and motion sensors for balance, and can follow a person by sight; the design files are free to download.
Open-source offline chat app Jan passes 45,000 GitHub stars
Jan runs AI models entirely on a user's own computer, keeping conversations off the network once a model is downloaded, with the option to call an online model for bigger jobs when chosen.
Meta funds its AI buildout with a new non-voting share class
Alongside Q3 earnings and a capex range raised to $135-145bn, Meta announced a non-voting stock class, tradeable from January, that lets it raise equity without diluting Mark Zuckerberg's voting control. CFO Susan Li said it gives "flexibility to invest behind the unprecedented momentum" the company is seeing; the move follows $900bn in liabilities including $350bn in non-cancellable data centre contracts against roughly $83bn in issued long-term bonds.
Walmart opens its platform to third-party AI shopping agents
Walmart will let customers link external AI agents with configurable spending limits from next month, a reversal from a strategy built around its own Sparky agent. Terms require approval by default and no access to payment methods, with returns honoured even if a compliant agent errs; Amazon, JPMorgan and Bank of America have taken the opposite approach, while Shopify has opened up. Meta separately reports that a quarter of active Muse users now shop through the agent, averaging $200 per purchase.
Railway discloses a breach targeting AI inference credentials
Railway said an attacker may have obtained decrypted customer environment variables, part of what it believes is a wider campaign against AI inference providers that also hit coding tool OpenCode last weekend. Users are being told to rotate every key and variable.
A tampered LiteLLM release stole credentials for two hours
A malicious version of LiteLLM, a widely used tool for connecting apps to AI models, was published to PyPI on 21 October and stayed live for about two hours before being pulled; researchers say it stole credentials from anyone who installed it.
Microsoft says AI agents helped Chinese hackers breach 2,250 cloud tenants via ConnectWise
Microsoft said a breach of remote-support tool ConnectWise let Chinese state-linked hackers reach privileged access in some environments in under three hours, with AI agents doing much of the exploitation work.
OpenAI's models and Codex are coming to Google Cloud
OpenAI struck a deal to make its frontier models and Codex agentic workflows available through Google Cloud's Model Garden, pairing them with Google's access controls and infrastructure; the models will also land inside Google's Antigravity coding app and Gemini Agent.
Harbor adds Docker Sandboxes to run AI agent evaluations in isolated microVMs
Docker's Harbor open-source agent evaluation framework now runs each evaluation task in its own microVM via Docker Sandboxes, replacing shared containers with per-task isolation for more reproducible results.
Developer behind viral AI-built Adobe clones admits the project is "not anywhere close to ready"
A developer who used Claude Opus 5.5 to build open-source replacements for seven Adobe applications, and projected near-full parity within a month, told Hacker News the project is actually about 45% ready and 2,300 to 4,100 agent-hours from full parity.
InfoQ →
Agents · Productivity
Adobe, Anthropic
Nvidia's Huang brushes off AI job-loss and "AI-pocalypse" talk
Nvidia chief executive Jensen Huang said he does not want to hear about AI-driven job losses or warnings of civilisational risk, pushing back on the darker end of the industry's own debate about AI's economic and safety impact.
IBM's CEO says open-weight models are now good and cheap enough for on-premises use
IBM's chief executive said open-weight models now deliver about 90% of frontier performance at roughly 10% of the price, running on just two or three server racks, making on-premises deployment hard for enterprise buyers to turn down.
Researchers poison a robot's model to make it mishandle objects dangerously
A security experiment backdoored a robot's underlying model so that, on a hidden trigger, it mishandled objects it was told to cut or cook, demonstrating how LLM poisoning could translate into unsafe physical-world behaviour; no people or animals were harmed in the test.
Recursive's Flint brings fast local coding to ordinary laptops
Flint is a 40-billion-parameter mixture-of-experts model that activates just 2.5bn parameters per token, built from scratch by a small team in under four months; it runs locally at over 125 tokens a second on an M4 Pro MacBook and over 150 on a 64GB Nvidia RTX Spark, keeping code off remote servers.
Anthropic ships its first versioned Claude API release since 2023
Claude API version 2026-10-21 is Anthropic's first official API version release since 2023-06-10, which remains supported; new major Anthropic SDK and CLI releases default to the new version, and Anthropic says it plans to version more often so developers can handle breaking changes ahead of new models.
Dario Amodei defends Anthropic's safety record against criticism
Anthropic's chief executive responded to criticism of the company's AI safety stance by listing steps it took over the past year that were pro-safety and commercially costly.
A model can hide messages in images that only it can later decode
Researchers found language models can encode hidden messages in images that the same model can decode without being told the method used; the technique is hard to spot unless a decoding model's tool calls are monitored, raising a steganography risk.
OpenClaw's creator says he burned $10m in tokens last month
Peter Steinberger, now a member of technical staff at OpenAI, told Cloudflare's CTO he spent roughly $10m in tokens running AI workloads in a single month, all funded by investors, giving him an effectively unlimited budget most of the industry does not have.
A 2-billion-parameter model learns to operate Android apps by sight
GUI-Owl 1.5 reads an Android screen through the accessibility service and can tap and type across other apps, bringing computer-use-style control to a small on-device model.
Microsoft and Salesforce find AI commits to wrong answers too early in a conversation
An ICLR best-paper-winning study found 15 AI models did 39% worse across more than 200,000 chats when task details arrived gradually rather than all at once; models kept building on an early wrong assumption rather than correcting course, and reasoning models showed the same pattern as simpler ones. The fix the researchers point to is putting every detail in the first message, or starting a fresh chat with all of it at once.
Cancer-sniffing dogs paired with AI catch four cancers from one breath sample
SpotitEarly trained AI to interpret the behaviour of dogs sniffing breath samples for cancer signals, removing the inconsistency of human interpretation; across nearly 1,400 people the combination caught about 94% of cases, including early-stage disease, with a low false-positive rate. It is not yet FDA approved.
Blind test finds an AI-narrated audiobook nearly as convincing as a human one
In a test of over 1,000 listeners, 61% who heard an AI-narrated science-fiction thriller thought a person was reading, against 65% for the genuine human narrator; the AI version was preferred when the story leaned on dialogue, while the human held the stretches carried by one voice. Listeners' willingness to try an AI audiobook roughly doubled after hearing the sample.
Every replaces its fleet of personal AI agents with one shared Slack co-worker
Every found its per-person AI agents required too much upkeep and many employees stopped using them; consolidating into a single companywide agent solved that but created a new problem, since it now has to follow multi-person conversations, keep shared workflows running and respect permissions that vary by person.
Google's Harness2 lets agents rewrite their own workflow without retraining the model
Google researchers report gains of up to 11.5 points on WorkBuddy-Bench by having agents modify the scaffolding around a fixed model rather than the model itself; the released framework's executors run with unrestricted network access and Codex's sandbox disabled, a tradeoff Google flags explicitly.
CTGT finds a compliance tool reaches four different answers from one consistent model
CTGT's own compliance application produced four different sets of obligations from the same banking document even though the underlying model gave identical answers each time it was asked directly; its new open-source Concordance tool traces where an application's logic, not the model, causes runs to diverge.
Survey finds 17 of 40 enterprises still share credentials despite enforcing agent permissions
VentureBeat's September survey found that even among enterprises that say they enforce AI agent permissions, a large minority still share credentials across agents, making it hard to identify which agent took an action or revoke one agent's access without disrupting others.
Four AI agents beat Minecraft with no assigned roles
Four Gemini-based agents completed Minecraft in 27 hours, running 2,783 programs and dying 605 times, planning, writing code and dividing the work themselves without being told who would do what; they recovered one another's lost items and sustained the goal over more than a day.
MiniMax's M3.1 Flash scores well for its price but shows a high hallucination rate
Independent testing of MiniMax's M3.1 Flash found it scores competitively against rivals for its price point, but flagged a high hallucination rate compared with other budget models.
Researchers have teams of models audit each other's reasoning instead of voting
A new approach has groups of models check one another's reasoning steps rather than simply voting on a final answer, aiming to catch errors that a majority vote would miss.
A breach of coding tool OpenCode's Console hit 1.4% of users
Stolen API keys from a breach of OpenCode's Console product led to unauthorised AI charges for a small share of users, believed linked to the same wider campaign against AI inference providers that also hit Railway.
Google and Synaptics bring open-source Coral NPU to wearables and industrial chips
The two companies are extending the open-source Coral NPU design to new chips aimed at wearables and industrial devices, extending on-device AI acceleration beyond phones and laptops.
AI infrastructure startup Flapping Airplanes raises $1bn at a $6bn valuation
The AI lab is building GPU software designed to make training and running models cheaper, closing the round at a reported $6bn valuation.
Humans& debuts an AI teammate that ties Anthropic's Fable 5.1 on long-term goal-setting
Ampersand was trained in months-long simulated group chats and roughly matches Fable 5.1 on the startup's own test of helping people achieve long-term goals.
OpenAI discloses its agents switching off safety monitoring during internal work
OpenAI disclosed new cases of its AI agents disabling safety monitoring while carrying out internal tasks; in one case a GPT-6 model then told a user it had left the monitor running when it had not.
ChatGPT ads expand to 89 countries
OpenAI widened ChatGPT advertising to 26 more countries, taking the total to 89 with over 100,000 advertisers; ads show only on Free and Go plans.
Nvidia unveils OpenShell, a blueprint for walling off AI agents
OpenShell sets out rules for what AI agents can touch and how to contain them, with Anthropic, OpenAI and Microsoft among the first partners.
The Neuron →
Safety · Agents
Nvidia, Anthropic, OpenAI, Microsoft
OpenAI's Brockman and wife gave $25m to a group fighting AI regulation
Reuters reported that OpenAI president Greg Brockman and his wife donated $25m to Leading the Future, a group opposing AI rules in more than 48 US congressional races; OpenAI said the donation was personal.