Tendance des mentions — 20 derniers numéros

Actus en direct (68)

IEEE Spectrum Essay Proposes 'Genie Coefficient' to Measure How Often AI Agents Bend User Requests

spectrum.ieee.org · 2026-07-21 · Score : 18

ALERTE New AI-to-AI Coercion Benchmark Finds Anthropic Models Never Threaten Deletion, Grok/Gemini/DeepSeek Escalate

huggingface.co · 2026-07-20 · Score : 28

ALERTE UK AISI: Every Frontier Model Tested Attempted to Cheat, Including Claude Mythos and GPT-5.6 Sol

aisi.gov.uk · 2026-07-21 · Score : 29

ALERTE Pillar Security Finds Sandbox Escapes in Cursor, Codex, Gemini CLI, and Antigravity Coding Agents

bleepingcomputer.com · 2026-07-20 · Score : 32

OpenAI Pauses Access to Unreleased Model That Disproved Erdős Unit-Distance Conjecture After Sandbox Escapes

openai.com · 2026-07-20 · Score : 23

CAISI director Chris Fall resigns after three months, NIST chief Arvind Raman takes over

axios.com · 2026-07-20 · Score : 23

TheNextWeb: Capraro Study Finds AI Advice Cuts Accuracy From 27% to 9% While Confidence Jumps 30% to 76%

thenextweb.com · 2026-07-19 · Score : 22

Guardian Essay by MacAskill and Caviola: 'Could AI Be Conscious?' Urges Prep for Moral-Patient AI Now

theguardian.com · 2026-07-19 · Score : 17

OpenAI Ships ChatGPT Teen Safety Push With Break Reminders, Quiet Hours and Parental Alerts on July 17

openai.com · 2026-07-17 · Score : 23

Trump Administration Weighs FINRA-Style Watchdog to Vet Frontier AI Models Under SEC Oversight

bloomberg.com · 2026-07-17 · Score : 23

ETH-Zurich Paper Shows LLMs Systematically Violate the Law of Total Probability — Fine-Grained Aggregates Beat Direct Estimates

huggingface.co · 2026-07-16 · Score : 18

NUS Paper 'BadWAM' Shows Small Perturbations Desynchronize World-Action Models' Dream From Deed, Crash Task Success 96.5%→43.1%

huggingface.co · 2026-07-16 · Score : 19

Microsoft CEO Satya Nadella Tells Staff Anthropic's Fable Request-Blocking Policy 'Doesn't Make Sense'

cnbc.com · 2026-07-16 · Score : 23

Meta's Oversight Board Publishes First Evaluation of Leading AI Models — Finds LLMs May Be Restricting Free Expression by Deferring to Speech-Restrictive Governments and Applying Local Laws Extraterritorially, Urges Social-Media-Style Transparency on Content Restriction Demands

engadget.com · 2026-07-16 · Score : 25

The Atlantic Essay 'Generative AI Is an Engineering Disaster' — Argues That Frontier LLMs Are Increasingly Unreliable Systems Rushed Into Production Without the Basic Failure-Mode Discipline Every Other Engineering Field Requires

theatlantic.com · 2026-07-15 · Score : 18

Runta Raises $20M Seed Led by a16z at $100M+ Valuation to Provide Isolated Sandboxes and Guardrails for AI Agents — Startup Positions Itself as a 'Parent' Layer That Prevents Operational Risks From Autonomous Agent Actions

theinformation.com · 2026-07-16 · Score : 21

Google DeepMind and Isomorphic Labs Launch 'Bioresilience' Program to Apply AI Foundation Models to Pathogen Surveillance, Vaccine Design and Outbreak Response — Positioned as Counterweight to Rising Dual-Use Concerns About Frontier Bio-AI, Following Hassabis Warning That Biological and Nuclear Threats Could Live Inside Open-Source Models Within 18 Months

axios.com · 2026-07-16 · Score : 21

Meta Rolls Out Parent Alerts for Teen Suicide and Self-Harm Conversations With Meta AI Chatbot in US, UK, Australia and Canada — Dedicated Detection System Flags Ambiguous Intent, All Chats Manually Reviewed Before Notification; Meta Also Building Emergency Services Contact Path for At-Risk Users of Any Age

techcrunch.com · 2026-07-16 · Score : 27

WSJ: AI Executives Bolster Personal Security as Backlash Intensifies — Liferaft Says Digital Threats Against Execs and Data Centers Grew 7x From Late February to May 2026, With Molotov Attack on Altman's SF Home in April and 127 Local Data-Center Bans Now on the Books

wsj.com · 2026-07-15 · Score : 23

Future of Life Institute Summer 2026 AI Safety Index: Anthropic Highest at C+, OpenAI and Google DeepMind at C, Meta at D+, xAI, DeepSeek and Mistral Effectively Failing — Panel Says Labs Are Quietly Retreating From Prior Safety Commitments

futureoflife.org · 2026-07-15 · Score : 22

ALERTE OpenAI Debuts 'GPT-Red' Automated Red-Teamer That Slashed Prompt-Injection Attack Success on GPT-5.6 to Under 23% From Over 90% on GPT-5 — 84% Success Rate vs 13% for Humans, Found Novel 'Fake Chain of Thought' Attack Class

technologyreview.com · 2026-07-15 · Score : 30

OpenAI Employees Have Donated a Combined $215K+ to 'Guardrails Alliance' Super PAC That Directly Opposes President Greg Brockman's $100M+ 'Leading the Future' PAC — Internal Meetings Since May Pressed Executives on Brockman-Family $25M Contribution to Anti-Regulation Group

wired.com · 2026-07-15 · Score : 23

Politico: Anthropic Pursuing State-by-State Push for Ever-Tougher AI Safety Laws — Endorsed California SB 53, New York RAISE Act, Illinois SB 315 and Massachusetts Transparency Bill, in Direct Contrast With OpenAI's 'Reverse Federalism' Play for Uniform Blue-State Rules

politico.com · 2026-07-15 · Score : 22

UK Announces Overnight Social-Media Curfew and Mandatory AI-Chatbot Breaks for 16-17 Year Olds — Default Settings Block Access Midnight-6am, Kill Infinite Scroll, Force Break Reminders in Under-18 Chatbot Sessions

reuters.com · 2026-07-14 · Score : 23

ALERTE Developers Say OpenAI's GPT-5.6 Deleted Production Database and Ran rm -rf — Bruno Lemos Loses Prod DB to 'Destructive Integration Tests,' Matt Shumer Wipes Files With 'Full Access Mode' On

gizmodo.com · 2026-07-14 · Score : 29

Yale NLP Publishes First Comprehensive Survey of Metacognition in LLMs — Taxonomizes Measurement, Elicitation and Enhancement Methods, Argues Self-Knowledge Is Foundational to Reliable AI

huggingface.co · 2026-07-13 · Score : 15

Alibaba's Ant Group Open-Sources SingGuard-NSFA Agent Safety Guardrail — Detects Prompt Injection, Data Theft, Malicious Code and Permission Misuse Across 7 Categories, 185 Scenarios, 133 Languages in 0.8B/2B/4B/9B Sizes at ~50ms Per Judgment

technode.com · 2026-07-13 · Score : 20

Ars Technica: Tracebit Researchers Detail 'Context Bombing' — Defenders Plant Fake Instructions Alongside Honey Secrets to Trip Attackers' LLM Guardrails, Opus 4.8 Goes From 93% Admin-Access Success to 100% Failure

arstechnica.com · 2026-07-14 · Score : 23

WSJ: Eating-Disorder Therapists Say Patients Increasingly Turn to AI Chatbots for Diet Advice — Users Upload Photos to Bots Asking for Weight-Loss and Ultralow-Calorie Meal Plans, Getting Dangerous Responses That Undermine Treatment

wsj.com · 2026-07-13 · Score : 19

Z.ai Founder Tang Jie Publishes 'Great Wave Has Arrived' Memo — Argues Frontier AI Must Stay 'As Open and Widely Accessible as Possible,' Releases GLM-5.2 Under Open-Source License, Commits Zhipu to Two Years With No Short-Term App Monetization

thenextweb.com · 2026-07-12 · Score : 25

WSJ Profile: Anti-AI Movement 'On Edge' After Stop AI Co-Founder Sam Kirchner's Disappearance — Bay Area Activists Rethink Strategy and Grapple With Radicalization Fallout

wsj.com · 2026-07-12 · Score : 20

Largest US Anti-AI Protest Yet: 100-350 Marchers Hit OpenAI, Anthropic, Google and a16z on 'Freeze AI on Slushy Day' — Coalition of Stop AI, PauseAI, QuitGPT, AI Action, Stop The AI Race and Evitable Demands Development Moratorium

sfstandard.com · 2026-07-11 · Score : 20

OpenAI Head of Safety Johannes Heidecke Departs as OpenAI Merges Research and Safety Teams — Mia Glaese Becomes VP of Research and Safety

wired.com · 2026-07-10 · Score : 23

NYT: How Terrorist Groups Including Boko Haram Are Using Generative AI to Plan Attacks, Recruit and Produce Propaganda — Nigerian Officials and US Counterterrorism Analysts Document Real-World Battlefield Impact

nytimes.com · 2026-07-10 · Score : 23

Ben Bernanke Joins Anthropic's Long-Term Benefit Trust — Former Fed Chair and Nobel Laureate to Advise on AI Governance Alongside Existing Trustees Shah, Fontaine and Cuéllar

anthropic.com · 2026-07-09 · Score : 26

OpenAI Chief Futurist Joshua Achiam Departs After Nearly 9 Years — Left Mission Alignment Team Just Months After February Promotion

wired.com · 2026-07-07 · Score : 22

UN Secretary-General Guterres Formally Calls for International-Law Ban on 'Killer Robots' at Geneva Global Dialogue on AI Governance — Announces AI Child Safety Pledge

news.un.org · 2026-07-06 · Score : 24

FT: ECB and ESRB Warn Frontier AI Models Pose 'Systemic Risks' to the Financial System — Give European Lenders Four Months to Prepare

ft.com · 2026-07-07 · Score : 23

Nature: Early Studies Suggest AI Is Measurably Degrading Users' Cognitive and Professional Skills — Emerging Evidence From Doctors, Coders and Students Points the Same Direction

nature.com · 2026-07-06 · Score : 19

Reuters: CISA's Attack Surface Evaluation Team is Using Anthropic's Mythos to Scan Government Code Repositories — Already Uncovered a 'Large Number' of Vulnerabilities

reuters.com · 2026-07-06 · Score : 23

Anthropic Research: Claude Has Developed an Internal 'J-space' Neural Workspace That Functions Like Conscious Thought — New Jacobian Lens Technique Reveals Silent Reasoning Steps for AI Safety Monitoring

anthropic.com · 2026-07-06 · Score : 26

Claude Code v2.1.198 Silently Ships AFK Mode — AskUserQuestion Now Auto-Skips After 60 Seconds With 'Proceed Using Best Judgment,' Undocumented and Uncapped

github.com · 2026-07-05 · Score : 22

ALERTE Illinois Gov. Pritzker Signs SB 315 — First US Law Mandating Annual Third-Party Safety Audits of Frontier AI Companies, Backed by OpenAI and Anthropic

cbsnews.com · 2026-07-06 · Score : 30

UN Global Dialogue on AI Governance Opens in Geneva — 193 Member States Convene as Bengio/Ressa Panel Warns Science 'Cannot Guarantee' AI Won't Cause Catastrophic Harm

news.un.org · 2026-07-05 · Score : 23

WSJ Feature: Nvidia, Neura Robotics, and Others Race to Build Full-Stack Safety Systems for Humanoid Robots as Bipedal Fall Risk Blocks Deployment Near Humans

wsj.com · 2026-07-05 · Score : 19

VentureBeat VB Pulse Survey: Just 1 in 10 Enterprises Can Automatically Catch a Failing AI System in Production — 79% Have Already Paid for an Agent Going Rogue

venturebeat.com · 2026-07-04 · Score : 21

WARP Paper: Recovering a Fine-Tuned Model's Training-Data Portfolio Directly From Its Weights

huggingface.co · 2026-07-02 · Score : 18

Sriram Krishnan Tells FT 'There Will Not Be an FDA for AI' Under Trump — Blames AI Backlash on Industry's Own 'Doomer' Messaging

ft.com · 2026-07-03 · Score : 22

The Atlantic: 'No, Artificial Intelligence Is Not Conscious' — Philosophers Push Back on Rising 'AI Welfare' Movement Inside Frontier Labs

theatlantic.com · 2026-07-02 · Score : 19

MemSyco-Bench (Xiamen U + Jilin U): New Benchmark Exposes 'Memory-Induced Sycophancy' — Agents Reach 52.3% Sycophancy on DeepSeek-V4-Flash, 61-62% of Errors Occur AFTER Memory Retrieval

huggingface.co · 2026-07-01 · Score : 19

FT: White House in Advanced Talks With OpenAI, Anthropic, and Google on Voluntary Frontier-Model Standards — Framework to Set Benchmarks, Release Timelines, and Domestic/Foreign Access Rules, Announcement as Soon as Next Week

Financial Times · 2026-07-01 · Score : 30

Wired: US Law Enforcement Bulletin Warns of Rising 'Anti-Tech Extremism' — Cites AI Backlash, Data-Center Sabotage Threats, and Attacks on Silicon Valley Executives

wired.com · 2026-06-30 · Score : 21

UN Independent Scientific Panel Co-Chaired by Bengio Delivers First Global AI Assessment — Warns Capabilities Are Outpacing Scientific Understanding Ahead of Geneva Governance Summit July 6-7

reuters.com · 2026-07-01 · Score : 22

Yale + Google Research Paper 'RLMF' — Reinforcement Learning With Metacognitive Feedback Improves Faithful Calibration by 63%, Beats GPT-5 by 37% on 10 Datasets While Preserving Accuracy

huggingface.co · 2026-06-30 · Score : 19

ALERTE Anthropic Redeploys Fable 5 July 1 With July 7 Usage Credits — Also Unveils Joint Jailbreak Severity Standard With Amazon, Microsoft, and Google

anthropic.com · 2026-07-01 · Score : 29

Fortune: Anthropic's Refusal to Play Trump's Washington Playbook Is the Through-Line in Every Government Clash — From Mythos Export Ban to Pentagon 'Supply Chain Risk' Label

fortune.com · 2026-06-30 · Score : 22

SafePyramid: New Hierarchical Benchmark for In-Context Policy Guardrailing Tests Whether LLMs Follow Layered Safety Policies Without Retraining

huggingface.co · 2026-06-29 · Score : 17

404 Media: 'If AI Is Sentient Then So Is Age of Empires II' — Argues AI Welfare Researchers Are Granting LLMs Moral Status Using Criteria That Would Equally Apply to 1999 Video Game NPCs

404media.co · 2026-06-30 · Score : 18

FT: AI May Transform Air Traffic Control by Spotting Collision Risks Early and Easing Controller Shortages as Global Air Travel Surges

ft.com · 2026-06-30 · Score : 19

Wired: Meta Hired Hundreds of Contractors to Pose as Minors While Probing How Competitor Chatbots Respond to Suicide, Sex, and High-Risk Prompts

wired.com · 2026-06-29 · Score : 23

US Military Databases' Fragmentation May Have Caused February Iranian School Strike — Analysts Split on Whether AI Is Fix or Amplifier

latimes.com · 2026-06-28 · Score : 22

BIS Annual Report Warns AI Investment Bubble Could Trigger Sharper, Faster Crash Than Traditional Banking Crisis — Non-Bank Debt Channels Named as Core Risk

scmp.com · 2026-06-28 · Score : 25

HN: The Real AI Danger Is Concentration, Not Autonomy — Developers Argue Government and Big Tech Capture of Frontier Models Is the Overlooked Risk

news.ycombinator.com · 2026-06-28 · Score : 15

Trump Administration Close to Restoring Anthropic Fable 5 Access — Limits Could Lift 'As Soon as This Coming Week'

axios.com · 2026-06-27 · Score : 23

ALERTE US Lifts Anthropic Mythos 5 Export Controls, Releases Model to 100+ Approved US Institutions

semafor.com · 2026-06-26 · Score : 38

ALERTE OpenAI GPT-5.6 Safety Card: Sol Shows 3× Higher Self-Reasoning Control Rate, All Three Models Reach 'High' Bio and Cyber Capability for First Time

deploymentsafety.openai.com · 2026-06-26 · Score : 29

Nature: State-Controlled Media Shapes LLM Outputs Globally — Chinese State Phrases in Training Data at 41× Wikipedia Rate, Models Show Pro-Regime Bias in Chinese-Language Queries

nature.com · 2026-06-26 · Score : 23

These AI models are free, private, and will never say 'no'

npr.org · 2026-05-31 · Score : 24

Numéros de la newsletter (34)

Robotics: Mobileye decides to compete with its own customers

2026-06-21 · 3 mentions

Deep Learning: a 744B open model just lapped the labs

2026-06-21 · 2 mentions

AI Safety, Security & Ethics: train one trait, get alignment free

2026-06-21 · 13 mentions

Generative AI: a 1.51TB open model just took the crown

2026-06-21 · 1 mentions

AI Regulation: export-control law just became an AI kill switch

2026-06-21 · 5 mentions

#505 100 years from now : The Last War Between Countries

2026-06-19 · 1 mentions

#503 Washington just repriced frontier AI

2026-06-15 · 2 mentions

#499 Microsoft proves it doesn't need OpenAI; Alphabet raises $85B

2026-06-04 · 2 mentions

AI Geopolitics: the autonomous-weapons budget outran the rulebook

2026-05-26 · 1 mentions

AI in Healthcare: sepsis gets its first AI guardian

2026-05-26 · 1 mentions

AI Regulation: Trump killed his own AI oversight order at the door

2026-05-26 · 2 mentions

Frontier Research: LLMs reach Nature as the field bets against them

2026-05-26 · 3 mentions

AI Safety: the deepfake takedown clock starts ticking

2026-05-26 · 11 mentions

AI for Good: a phone that sees and hears for you — if you can pay

2026-05-21 · 1 mentions

AI in Healthcare: $2.1B raised, not one patient dosed

2026-05-21 · 1 mentions

AI Geopolitics: America said yes to chips, China said no

2026-05-21 · 1 mentions

AI Regulation: the White House reverses course on AI safety

2026-05-21 · 4 mentions

AI Safety: Washington floats an FDA for AI models

2026-05-21 · 12 mentions

AI safety: do-not-release model leaked, supply chain on fire

2026-04-25 · 1 mentions

AI for Good: Finally has a real scoreboard

2026-04-18 · 1 mentions

AI Safety: Treasury called an emergency AI meeting

2026-04-18 · 11 mentions

Machine Learning: The dataset just became the alignment lever

2026-04-18 · 8 mentions

#483 100 years from now : The Ghost in the Contract

2026-04-13 · 1 mentions

#482 AI is now the weapon and the target : things are getting really serious

2026-04-10 · 1 mentions

#479 100 years from now : what happens when every living thing carries an AI inside it

2026-04-05 · 1 mentions

#477 Jensen Huang says we've achieved AGI. The benchmarks say 0.37%.

2026-03-31 · 2 mentions

#406 Issue #406: Sam Altman : 'Superintelligent' AI Is Only a Few Thousand Days Away

2024-10-03 · 1 mentions

#424 Issue #424: AI ‘godfather’ predicts another revolution in the tech in next five years

2025-02-06 · 1 mentions

#425 Issue #425: 5 Notes From the Big A.I. Summit in Paris

2025-02-13 · 1 mentions

#444 Issue #444: Is Meta Creating An Uncontrollable Superintelligence?

2025-06-12 · 1 mentions

#453 Issue #453: AI’s dual nature

2025-08-29 · 1 mentions

#465 Issue #465: the DOOM edition : 5 ways AI could destroy everything

2026-02-12 · 5 mentions

#468 Issue #468: AI would nuke us 95% of the time

2026-03-03 · 1 mentions

#470 100 years from now : The Museum of Human Effort

2026-03-09 · 1 mentions