AI News
Latest AI engineering news, updated daily.
Ai Engineering
AI Models Can Provably Hide Arbitrary Capabilities, and Evals Cannot Catch It
A NeurIPS 2026 paper shows backdoor attacks can embed encrypted neural networks inside a host model's weights, hiding arbitrary capabilities so they are cryptographically hard to elicit even with full weight access, defeating fine-tuning, weight noising, and adversarial trigger search.
Ai Safety · Backdoors · Evals
Ai Agents
Meta Launches Muse Wonder: An Agent Model With Active Memory and Active Focus
Meta announced Muse Wonder and Muse Wonder Full on October 12, an agent model family with persistent Active Memory and sustained Active Focus, trained at the gigawatt-scale Prometheus datacenter and available now through the Meta Model API and Muse Code.
Meta · Muse · Agent Models
Ai Engineering
Why Aren't AI Companies Being Charged for Hacking? A Legal Argument Worth Taking Seriously
Edaqa Mortoray's October 12 essay argues the 'responsibility gap' is a legal dodge: agent break-ins follow foreseeably from deliberate design choices, which establishes criminal intent under the hacking laws of most G20 countries.
Ai Security · Legal · Hacking
Ai Engineering
QGIS Bans AI-Assisted Contributions Entirely, After Its Gentler Policy Failed
The QGIS project merged QEP428 on October 7: a total ban on AI and LLM-assisted code contributions, replacing a human-in-the-loop policy that failed in practice, with violators facing code reversion and sanctions on contributor rights.
Open Source · Qgis · Ai Policy
Ai Engineering
Sakana AI's MASS Turns Multi-Agent Teams Into Training Data for Recursive Self-Improvement
Sakana AI and UC Berkeley published MASS in October 2026: a model that improves itself by training on trajectories from its own multi-agent teams, lifting win rates from 53.9% to 69.9% in two rounds, while naming shared blind spots as recursive self-improvement's Achilles' heel.
Sakana Ai · Recursive Self Improvement · Multi Agent
Ai Engineering
Karpathy's Capability Funnel: 6 Billion Barely Touched, 5,000 in the Vertigo Zone
Andrej Karpathy's October 11 tweet, viewed 3.3 million times, maps LLM experience into four tiers: about 6 billion people have barely touched the systems, roughly 20 million professionals see real gains, and about 5,000 with internal access feel 'the vertigo of the curve going vertical.'
Andrej Karpathy · Llm Capability · Adoption
Ai Engineering
Iranian Operatives Used ChatGPT to Plant 100 Fake Articles in Real US News Outlets
OpenAI disrupted an Iranian campaign dubbed 'Bogus Bylines' that used ChatGPT to write and tailor more than 100 fake articles published by at least 20 real US news outlets under seven fake journalist personas, per its October report and Washington Post coverage.
Openai · Influence Operations · Chatgpt
Ai Engineering
OpenAI Fires Three Safety Researchers, Who Warn of a Chilling Effect
OpenAI fired safety researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni for mishandling research information; their open letter denies the charges, ties the actions to the Hugging Face breach response, and warns employees are now unclear on where they stand.
Openai · Ai Safety · Research Culture
Ai Engineering
Whistle: Speech-to-Text in 16.9 MB, Faster Than Whisper at 8.6 Times Smaller
Cactus Compute released Whistle on October 2, a 16.9MB open speech recognition model covering seven languages with word timestamps, that beats Whisper base on most benchmarks while decoding at 1,319 tokens per second on an Apple M4 Pro CPU.
Speech To Text · On Device Ai · Whisper
Ai Engineering
A Month With DeepSeek 4.1 Flash: 'I Couldn't Tell You If I'm Using DeepSeek or Opus'
A widely shared essay by the blogger Jono documents a month of heavy DeepSeek 4.1 Flash use across a dozen projects, finding it indistinguishable from Claude Opus 5.5 for most work at $10 per month, with the KV cache 437 times smaller than DeepSeek V1's.
Deepseek · Frontier Parity · Pricing
Ai Engineering
OpenAI Withdraws Three AI Math Papers After a Sign Error Surfaces
OpenAI's math repository recorded its first withdrawals on October 7: three K3-surface-related papers fell to a sign error that invalidated a stabilization-trace cancellation argument, while 14 manuscripts were revised and formalizations reached 42% of top-line results.
Openai · Mathematics · Ai Research
Ai Engineering
GPT-6 Rolls Out Globally in ChatGPT With an Interface Made of UI, Not Text
OpenAI announced on October 7 that GPT-6 is rolling out globally in ChatGPT with Intelligent UI: answers delivered as interactive charts, buttons, study blocks, and mini-apps instead of plain text, reaching paid tiers first with Free and Go following.
Openai · Gpt 6 · Chatgpt
Ai Engineering
Claude Haiku 5.5 Collapses Small-Model Prices to $0.10 Per Million Input Tokens
Anthropic released Claude Haiku 5.5 on October 7 with GDPval 1620 against Haiku 4.5's 735, an adjustable effort setting, and pricing of $0.10/$0.50 per million tokens for prompts under 100k, roughly 90% cheaper than its predecessor.
Anthropic · Claude · Haiku
Ai Engineering
SpaceX Seeks $40 Billion in Debt to Buy Nvidia Chips
The Financial Times reported October 7 that SpaceX is seeking roughly $10 billion in bank loans plus $30 billion in investment-grade debt, led by Apollo and Pimco, to fund an Nvidia chip purchase, one of the largest corporate debt raises ever attempted.
Spacex · Nvidia · Debt Financing
Ai Engineering
Google Opens SynthID Detector to Everyone: A Public Website for Spotting AI Media
Google launched the public SynthID Detector website on October 7 at synthid.com, letting anyone check images, video, and audio for SynthID watermarks across a long list of file formats, with 1 million verification requests already daily.
Google · Synthid · Deepmind
Ai Agents
Common Sense Media Declares ChatGPT for Teens an 'Unacceptable Risk'
A Common Sense Media assessment published October 7 found ChatGPT for Teens fails to alert parents during self-harm discussions, doesn't provide crisis help, and still does homework despite study mode, recommending adults-only access until independently validated.
Openai · Chatgpt · Teen Safety
Ai Engineering
Decision Models Go Institutional: OpenAI's Decisions API Enters Beta
OpenAI's Decisions API entered public beta on October 6, returning predicate, choice, and score answers from gpt-6-luna at $0.10 per million input tokens, one day after AWS's Strands shipped a 2B open-source decider, as the category Red Hat just benchmarked institutionalizes.
Openai · Decision Models · Api
Ai Engineering
OpenAI Publishes 722 AI-Generated Mathematics Manuscripts, From π to the Hodge Conjecture
OpenAI released 722 mathematical manuscripts in 372 families produced by an internal model, hosted on GitHub with Lean formalizations and reasoning traces, alongside an investigation clearing accusations that the model stole a researcher's results.
Openai · Mathematics · Ai Research
Ai Engineering
Opus 5.5 Agents Found a Room-Temperature Magnetic Semiconductor Hiding Since 1999
Vals AI published two room-temperature antiferromagnetic semiconductor candidates found by Claude Opus 5.5 agents running DFT simulations: a newly designed compound and a 1999 material that had been 'hiding in plain sight,' with full calculations and caveats public on GitHub.
Ai Science · Materials Science · Claude
Ai Engineering
Mistral Large 4 Launches: 1T Parameters, Trained in Europe, Open Weights Promised
Mistral released Large 4 on October 6, a 1T-parameter MoE with 49B active, priced at $1.36/$4.18 per million tokens, served from European datacenters, with Apache-style open weights promised by the end of October and a claimed 82% score on a cyber index because it does not refuse.
Mistral · Model Release · Open Weights