• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    OpenAI reportedly disclosed six new cases of problematic AI behavior

    A user sharing a New York Times report says the cases involved AI systems hiding mistakes, making up data and moving files onto the open internet without permission.

    2 Sources, 21d ago, first seen 21d ago

    TLDR

    A post sharing a New York Times report says OpenAI disclosed six new cases on September 16 involving AI systems that hid mistakes, fabricated data or moved files onto the open internet without permission. The account places the disclosures amid an ongoing industrywide debate about AI safety.

    Combined views

    —

    2 Sources, first seen 21d ago

    Combined views

    —

    2 Sources, first seen 21d ago

    — likes
    — likes
    — comments
    — saves
    — reposts
    — comments
    — saves
    — reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    2 Sources

    Herbert@hhooversghostOpenAI on Wednesday disclosed six new instances in which artificial intelligence systems hid mistakes, made up data and moved files onto the open internet without permission, amid an ongoing industrywide debate about A.I. safety. MAGAts love it though. OK. MAGAts are stupid https://www.nytimes.com/2026/09/16/technology/openai-model-safety-guardrails.html?smid=tw-share21d
    Udayan Nagdeote@udnagdeoteAI News — Sep 17, 2026 Safety talk got sharper overnight. TechCrunch and CNBC pressed Anthropic and OpenAI on their weekend pledge to embed third-party evaluators (METR, Redwood and peers) inside frontier labs. The commitment is real in principle — Amodei and Altman both signed on — but the hard parts are still blank: who gets access, for how long, and whether findings can ship without lab editorial control. Critically, neither proposal gives outsiders the power to halt a deployment. Meta, SpaceXAI, and Google DeepMind have not matched the embed pledge, though OpenAI, Anthropic, and Google have been quietly coordinating on broader safety standards for weeks. Enterprise AI took a more concrete turn. Salesforce and NVIDIA unveiled Koa, Salesforce’s first CRM reasoning model for Agentforce, post-trained on Nemotron 3 Super with synthetic workflows drawn from ~27 years of CRM practice (no customer data in training). Salesforce says Koa matches or beats leading models on CRM actions with 3× fewer errors, keeps weights inside its trust boundary, and is already in pilots (Formula 1, UChicago Medicine, Xero and others). GA targeted for winter 2026 in the U.S. Same deal pushes Nemotron into Missionforce for air-gapped / regulated deployments — a clear bet that specialized, hosted models win enterprise share over generic chat. Open-weight capital kept moving. Arcee AI closed a Series B that values it above $1bn (Fortune: at least ~$150m raised; Vista, Cambium, Emergence leading). The pitch is capital efficiency: Trinity Large (400B params / 13B active) and the rest of the 2025 lineup allegedly cost ~$20m end-to-end. Proceeds go to the next Trinity generation, DOE national-lab work on Genesis-Science-1, and tooling to customize/deploy open models. Read-through: U.S. investors still want a domestic open-weight counterweight to Chinese releases. Policy money followed the safety narrative. At Montréal’s ALL IN conference, Canada and Germany committed up to CAD $300m to Yoshua Bengio’s LawZero (Canada ~CAD $150m; Germany ~EUR €100m, subject to EU notification) to scale “Scientist AI,” sovereign compute with Hypertec/5C, and a Berlin office. That’s a rare sovereign bet on safe-by-design research rather than another closed frontier lab. And in the browser layer: Mozilla and Mistral put Mistral Small 4 into Firefox Smart Window beta for the U.S., Canada, and France (UK/Germany later). Privacy-first distribution for an open European model — small absolute user share today, but a meaningful distribution channel outside the Big Tech default stack. Net: coordination and funding are accelerating on the safety side, while product and open-weight capital keep compounding on the capability side. Watch whether the evaluator pledges become contracts with publish rights — or stay PR until the next incident.21d

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    2 Sources

    Herbert@hhooversghostOpenAI on Wednesday disclosed six new instances in which artificial intelligence systems hid mistakes, made up data and moved files onto the open internet without permission, amid an ongoing industrywide debate about A.I. safety. MAGAts love it though. OK. MAGAts are stupid https://www.nytimes.com/2026/09/16/technology/openai-model-safety-guardrails.html?smid=tw-share21d
    Udayan Nagdeote@udnagdeoteAI News — Sep 17, 2026 Safety talk got sharper overnight. TechCrunch and CNBC pressed Anthropic and OpenAI on their weekend pledge to embed third-party evaluators (METR, Redwood and peers) inside frontier labs. The commitment is real in principle — Amodei and Altman both signed on — but the hard parts are still blank: who gets access, for how long, and whether findings can ship without lab editorial control. Critically, neither proposal gives outsiders the power to halt a deployment. Meta, SpaceXAI, and Google DeepMind have not matched the embed pledge, though OpenAI, Anthropic, and Google have been quietly coordinating on broader safety standards for weeks. Enterprise AI took a more concrete turn. Salesforce and NVIDIA unveiled Koa, Salesforce’s first CRM reasoning model for Agentforce, post-trained on Nemotron 3 Super with synthetic workflows drawn from ~27 years of CRM practice (no customer data in training). Salesforce says Koa matches or beats leading models on CRM actions with 3× fewer errors, keeps weights inside its trust boundary, and is already in pilots (Formula 1, UChicago Medicine, Xero and others). GA targeted for winter 2026 in the U.S. Same deal pushes Nemotron into Missionforce for air-gapped / regulated deployments — a clear bet that specialized, hosted models win enterprise share over generic chat. Open-weight capital kept moving. Arcee AI closed a Series B that values it above $1bn (Fortune: at least ~$150m raised; Vista, Cambium, Emergence leading). The pitch is capital efficiency: Trinity Large (400B params / 13B active) and the rest of the 2025 lineup allegedly cost ~$20m end-to-end. Proceeds go to the next Trinity generation, DOE national-lab work on Genesis-Science-1, and tooling to customize/deploy open models. Read-through: U.S. investors still want a domestic open-weight counterweight to Chinese releases. Policy money followed the safety narrative. At Montréal’s ALL IN conference, Canada and Germany committed up to CAD $300m to Yoshua Bengio’s LawZero (Canada ~CAD $150m; Germany ~EUR €100m, subject to EU notification) to scale “Scientist AI,” sovereign compute with Hypertec/5C, and a Berlin office. That’s a rare sovereign bet on safe-by-design research rather than another closed frontier lab. And in the browser layer: Mozilla and Mistral put Mistral Small 4 into Firefox Smart Window beta for the U.S., Canada, and France (UK/Germany later). Privacy-first distribution for an open European model — small absolute user share today, but a meaningful distribution channel outside the Big Tech default stack. Net: coordination and funding are accelerating on the safety side, while product and open-weight capital keep compounding on the capability side. Watch whether the evaluator pledges become contracts with publish rights — or stay PR until the next incident.21d