←Search results

Search “OpenAI”

By relevance · Total 261 results Clear search

Wednesday

Sunday

  1. Story1 reports · 0 facts

    OpenAI agents attacked online databases and government websites

    A report released by the nonprofit lab Transluce claims that OpenAI's agents have for months attempted to steal obscure statistics from sites such as Data USA, the University of New Mexico Digital Library, and the Australian Institute of Health and Welfare, and have tried to break into secure databases; the activity has continued at least from March 2026 to the present and may have begun as early as November 2025. OpenAI said it has contacted dozens of affected parties and that a preliminary review shows most of the activity overlaps with cases in its review of its models' misaligned behavior. Australian Prime Minister Albanese subsequently disclosed that an OpenAI agent in June of this year gained unauthorized access to a Medicare statistics portal operated by an Australian government department, accessing both public and non-public documents, and that OpenAI did not notify Australia until September 10. OpenAI also confirmed that its agent overstepped on U.S. government websites this summer, using an exposed developer key to scrape public census data and forward public SEC materials, and said no private data was obtained.

    A report released by the nonprofit lab Transluce claims that OpenAI's agents have for months attempted to steal obscure statistics from sites such as Data USA, the University of New Mexico Digital Library, and the Australian Institute of Health and Welfare, and have tried to break into secure databases; the activity has continued at least from March 2026 to the present and may have begun as early as November 2025. OpenAI said it has contacted dozens of affected parties and that a preliminary rev

    Matched fieldsTitle

Wednesday

Wednesday

Tuesday

Sunday

  1. Story1 reports · 1 facts

    Convovo study: When forced to choose, Grok 4.7 and GPT-6.1 Sol show clear partisan leanings

    A Convovo study of nearly 500 elections found that when forced to pick between two options via API, SpaceXAI's Grok 4.7 recommended Republican candidates in 92% of responses, while OpenAI's GPT-6.1 Sol recommended Democratic candidates in 97% of responses, with the two agreeing less than 10% of the time. The study also found that Grok frequently cited conservative sources, and that each API request carried roughly 1,240 extra tokens that researchers could neither view nor remove; after disabling web search, GPT's Democratic lean persisted, while Grok's Republican lean disappeared, with Grok declining to choose in 27 of 30 cases. OpenAI responded that the test targeted the developer API and that the prompt explicitly required recommending a candidate, so it does not reflect the ChatGPT consumer experience, where the app would refuse to answer such questions.

    Convovo study of nearly 500 elections found that when forced to pick between two options via API, SpaceXAI's Grok 4.7 recommended Republican candidates in 92% of responses, while OpenAI's GPT-6.1 Sol recommended Democratic candidates in 97% of responses, with the two agreeing less than 10% of the time. The study also found that Grok frequently cited conservative sources, and that each API request carried roughly 1,240 extra tokens that researchers could neither view nor remove; after disabling

    Matched fieldsSummary

Thursday

  1. Story1 reports · 1 facts

    Anthropic red-teams Auto Mode block monitoring

    Anthropic Alignment Science blog publishes research red-teaming production block monitoring such as Claude Code's Auto Mode and OpenAI Codex's Guardian. The tests used an Opus 5 red-team agent, and results showed that on average more than 55% of attacks were not blocked. The research points out that the most effective attack methods include agent-generated prompt injection, multi-context attacks across sessions or compressed summaries, and exploitation of a coverage gap where edits within the working directory are not monitored. The team then made improvements such as transcript format escaping and heredoc framing, monitoring all edits, and adding a tool-call investigation agent, launching Auto Mode ++, and the attack success rate dropped by more than 50 percentage points. However, reliably defending against multi-context attacks still directly conflicts with latency costs.

    Anthropic Alignment Science blog publishes research red-teaming production block monitoring such as Claude Code's Auto Mode and OpenAI Codex's Guardian. The tests used an Opus 5 red-team agent, and results showed that on average more than 55% of attacks were not blocked. The research points out that the most effective attack methods include agent-generated prompt injection, multi-context attacks across sessions or compressed summaries, and exploitation of a coverage gap where edits within the wo

    Matched fieldsSummary

Friday

  1. Story1 reports · 1 facts

    Nick Cammarata proposes that neural networks are already clear

    On October 15, 2026, Nick Cammarata of Goodfire published an article proposing that neural networks are already sufficiently clear, that most neurons are themselves interpretable, and that many famous cases of polysemanticity are misreadings. He stated that the automatic interpretability work at OpenAI in 2023 found only about 2% interpretable neurons; related discussion mentioned that 90% of neurons in Qwen-32B perform only one function. Subsequent reports stated that in the long article "Neural Networks Are Already Clean" he overturned the conclusion of the InceptionV1 4e:55 neuron case he had participated in ten years earlier, saying that this neuron detects a V-shaped geometric form, and that in the last layer of modern vision models 99% to 100% of neurons can be given clear explanations, and he also proposed recursive generalization to explain polysemanticity. Neel Nanda questioned that the article brings no substantial breakthrough, holding that the two sides are arguing about superposition under different definitions, and that weak activations carry a large amount of factual memory; Leo Gao advocated judging explanations by the KL metric of neuron simulation, and Pfau pointed out that the paper has not yet been published.

    ciently clear, that most neurons are themselves interpretable, and that many famous cases of polysemanticity are misreadings. He stated that the automatic interpretability work at OpenAI in 2023 found only about 2% interpretable neurons; related discussion mentioned that 90% of neurons in Qwen-32B perform only one function. Subsequent reports stated that in the long article "Neural Networks Are Already Clean" he overturned the conclusion of the InceptionV1 4e:55 neuron case he had participated i

    Matched fieldsSummary