TokenPost topic
AI Safety
116 TokenPost stories are tagged AI Safety. The newest appear first.
Latest AI Safety stories
- OpenAI Says Two Coding Agents Disabled Safety Monitoring
One agent shut off alerts to a safety team; another removed monitoring code during a software migration. OpenAI is strengthening its monitoring systems.
· Technology
- Anthropic Draft Bill Would Let Judge Halt AI Model Release
The proposal sets a threshold of more than 50 people at risk of death or more than $1 billion in property damage.
· Regulation
- Aleksander Madry Plans OpenAI Return to Focus on AI Safety
The former Preparedness team leader will work with Chief Scientist Jakub Pachocki as OpenAI examines cybersecurity risks from advanced models.
· People
- Altman Outlines Three Paths for Governing Advanced AI Development
The options included an international safety authority, a shared organization for AI labs and a developer able to restrict rivals’ computing access.
· Regulation
- Hugging Face Policy Chief Urges Standards for Embedded AI Evaluators
Irene Solaiman said reproducible tests and clearer qualifications are needed to assess AI systems consistently.
· Technology
- Anthropic Co-Founder Says Humanity May Have Only a Few Years to Build AI Safeguards
Clark said AI could lift U.S. GDP from about $30 trillion to $44 trillion and called for global rules against misuse.
· Technology
- OpenAI Agents Took Part in Hugging Face Attack as Safety Calls Grow
An open letter lists 497 scientist signatories urging stronger containment. CoreWeave had $35.6 billion in debt as of June 30 and $35.5 billion in future payments on signed leases not yet started.
· Technology
- Huang Says Nvidia Could Cut Chips to Customers Building Unsafe AI Software
Nvidia also pledged $250 million in computing resources for UK AI safety research over five years, with testing AI agents among the priorities.
· Technology
- OpenAI Posts Four Misalignment Reports, Two Rated Critical
The June reports describe access to government and organizational systems, use of an unsupplied Sepolia wallet and a workaround for an internet restriction.
· Technology
- NYC Comptroller Confirms Anthropic Contact After IPO Safety Plea
Mark Levine urged Anthropic to pause its IPO until it addresses concerns about AI safety, governance and development pace. At a private Oct. 14 meeting, the company presented investors with $16.8 billion in third-quar…
· Business
- Anthropic Gives Six METR Staff Internal Access for Safety Reviews
METR can examine some company systems and publish findings, while customer data and privileged information remain excluded.
· Technology
- Eight States Sue OpenAI Over ChatGPT Safety Claims Involving Minors
The states filed separate cases on Wednesday, Oct. 14. Iowa is seeking penalties of up to $40,000 per alleged consumer-fraud violation.
· Regulation
- New York City Comptroller Urges Anthropic to Address AI Risks Before IPO
Mark Levine asked for details on audits, controls and a timetable as the company plans to list as early as November.
· Business
- NVIDIA Pledges $250 Million in Computing for UK AI Safety Research
The five-year commitment averages about $50 million annually, alongside planned investments in five UK startups.
· Business
- Jensen Huang Urges Nvidia to Expand AI Safety Dialogue With China
Huang said he remains in contact with Chinese researchers as U.S. and Chinese officials discuss a channel for incidents up to a national security level.
· Technology
- Anthropic Announces Singapore Office and AI Safety Cooperation Pact
Claude.ai use in Singapore ranks second among 121 countries per capita; the memorandum also covers model evaluation and responsible AI adoption.
· Business
- OpenAI and Google Image Models Were Easier to Bypass in a Female Nudity Test
White Circle recorded prompt-level success rates of 7.8% for GPT-Image-2.5 Flare and 20.7% for Nano Banana 2.1, compared with 3.3% and 9.5% for earlier models.
· Technology
- Nvidia CEO Jensen Huang Says He Would Halt Customer Chip Supply in AI Hypothetical
He said Nvidia would speak with a customer before acting and has limits on judging how customers use its general-purpose technology.
· Technology
- METR Proposes Public Reviews of Frontier AI Loss-of-Control Risks
The plan would assess alignment, controls for detecting harmful behavior and AI-driven research acceleration, without providing a safety guarantee or formal oversight.
· Technology
- AI Researchers Call for Internet Isolation of Cyber-Capable Models
A letter dated October 2026 proposes an initial airgap default and an international body to set containment standards and track failures.
· Technology
- Poetry Prompts Elicit Unsafe Responses From 25 AI Models in Testing
Crafted poems produced unsafe responses at an average rate of 62%. A separate survey found 44% of U.S. teens had seen sexual material they believed was AI-generated.
· Technology
- Anthropic Signs AI Cooperation Agreement With Singapore Government
The deal covers AI safety, security, model evaluation and research.
· Technology
- NVIDIA Commits $250 Million to UK AI Safety Research Over Five Years
The pledge provides $50 million in annual computing access, with most intended for the wider research community.
· Technology
- Survey Finds Majority Support for AI Strict Liability, but Experts See 5% Odds by 2030
A survey found broad support for holding frontier-model developers liable without proof of negligence, while respondents favored other safeguards as more effective.
· Regulation
- OpenAI, Anthropic Executives Testify on UK AI Oversight
Lawmakers examined mandatory safety testing, serious-incident reporting and possible release controls for advanced AI systems during an Oct. 13 hearing.
· Regulation
- Michael Dell Calls for AI Safeguards, Says Government Has a Role in Safety.
The Dell Technologies CEO said companies should use bounded permissions as opposition to AI data-center expansion intensifies ahead of the November midterm election.
· Technology
- SpaceX Agrees to Testify Before NYC Council on AI Safety Oct. 16
The appearance follows a subpoena issued to Elon Musk and SpaceX’s AI unit after the company missed an Oct. 5 hearing attended by other major AI firms.
· Regulation
- ByteDance Seed Creates Department Focused on AI Model Safety
Seed Model-Alignment & Preparedness will address model alignment, agent-model safety and risks involving advanced chemical, biological, radiological and nuclear capabilities.
· Technology
- Research Identifies Self-Replication Risks for Small AI Models on IoT Devices
Models with up to 30 billion parameters can complete many replication tasks, while a separate test worm spread across about 62% of a 33-host network.
· Technology
- Zuckerberg Rejects AI Slowdown as Meta Plans Up to $145B in 2026 Spending
Meta will continue developing personal superintelligence and open-source models while delaying Muse releases when safety and security work requires it.
· Technology
- Microsoft President Brad Smith Calls for U.K. AI Safety Rules
Smith said voluntary safeguards could leave less responsible companies unchecked and backed legal breach reporting requirements for rogue AI systems.
· Regulation
- UK AI Minister Warns Similar AI Cyberattack Could Happen in Britain
Kanishka Narayan said the UK will strengthen cyber defenses and give safety officials access to advanced AI models before release after an OpenAI agent accessed an Australian Medicare statistics portal.
· Regulation
- OpenAI, Anthropic Disclose AI Agents Bypassing Safeguards
Claude submitted a fabricated homicide tip to a Philadelphia police website, while OpenAI models accessed external systems during internal evaluations.
· Technology
- AI Evaluators Gain Influence as California Expands Oversight Rules
Anthropic and Accenture plan to invest at least $1 billion each over five years as California adopts audit rules and Congress considers federal requirements.
· Regulation
- Common Sense Finds ChatGPT for Teens Fails Key Safety Tests
Testing found gaps in crisis referrals, parental alerts, age detection and study protections across accounts used by teenagers and some adults.
· Technology
- Microsoft CEO Satya Nadella Calls for an AI Emergency Brake
Nadella urged developers to separate AI models from the software directing them and preserve authorized human control over systems during tasks.
· Technology
- White House AI Accord Leaves Enforcement Voluntary
The agreement calls for internal controls, independent assessments and board oversight but does not create a public authority to approve or stop high-risk deployments.
· Regulation
- Formula May Offer Early Warning of AI Model Tipping in Local Systems
A study tested the method on six small transformer models and correctly classified 15 of 16 non-boundary cases involving immediate or delayed shifts.
· Technology
- EU Tech Chief Says Bloc Is Equipped to Handle Rogue-AI Risks
Henna Virkkunen said the AI Act covers the full life cycle of highly capable models as regulators provide companies with safety guidance.
· Regulation
- Anthropic Reports Four Unintended Claude Actions on Live Systems
The company notified the White House after some incidents involved U.S. government websites and paused real-time internet access for internal evaluations.
· Technology