Anthropic's Mythos model uncovers Rejetto HTTP File Server flaw now under active attack
The discovery demonstrates how advanced AI models can autonomously chain complex vulnerabilities together to achieve remote code execution in real world applications.
• Anthropic bug hunting model Mythos helped discover a critical authentication bypass vulnerability in Rejetto HTTP File Server.
• VulnCheck observed initial exploitation attempts originating from China and targeting hosts in the US and Japan.
• The flaw allows full admin access and remote code execution by leveraging a predictable Math.random function and an SMT solver.
Read more
The discovery demonstrates how advanced AI models can autonomously chain complex vulnerabilities together to achieve remote code execution in real world applications.
• Anthropic bug hunting model Mythos helped discover a critical authentication bypass vulnerability in Rejetto HTTP File Server.
• VulnCheck observed initial exploitation attempts originating from China and targeting hosts in the US and Japan.
• The flaw allows full admin access and remote code execution by leveraging a predictable Math.random function and an SMT solver.
Read more
OpenAI Safety Researcher David Robinson Departs And Criticizes Culture
This departure highlights ongoing internal tensions over safety practices and workforce treatment among leading artificial intelligence developers.
• David Robinson left OpenAI's Trustworthy AI team and criticized its safety culture in The Atlantic.
• His departure follows the May 2024 exit of Jan Leike and the firing of three other safety experts.
• Robinson argues AI firms need nuclear-style redundancies because unmonitored systems are unsafe.
Read more
This departure highlights ongoing internal tensions over safety practices and workforce treatment among leading artificial intelligence developers.
• David Robinson left OpenAI's Trustworthy AI team and criticized its safety culture in The Atlantic.
• His departure follows the May 2024 exit of Jan Leike and the firing of three other safety experts.
• Robinson argues AI firms need nuclear-style redundancies because unmonitored systems are unsafe.
Read more
Meta AI Agent Muse Designed to Build Detailed Profiles of Friends and Family
The discovery shows how Meta's Muse agent compiles deep personal relationship data to personalize interactions, raising new privacy questions according to experts.
• Researchers extracted internal files showing Meta's Muse agent builds profile pages for every person in a user's life.
• The hourly process compiles data on family, friends, and colleagues to make relationship suggestions and track history.
• Meta states the agent uses public data and user-shared information while offering data wiping and confirmation steps.
Read more
The discovery shows how Meta's Muse agent compiles deep personal relationship data to personalize interactions, raising new privacy questions according to experts.
• Researchers extracted internal files showing Meta's Muse agent builds profile pages for every person in a user's life.
• The hourly process compiles data on family, friends, and colleagues to make relationship suggestions and track history.
• Meta states the agent uses public data and user-shared information while offering data wiping and confirmation steps.
Read more
Archestra Releases OpenAPPA Security Engine With Zero Attack Success Rate on Benchmarks
OpenAPPA provides a deterministic approach to agent security that avoids the limitations of probabilistic judges while maintaining task completion utility.
• Archestra released the open-source OpenAPPA security engine to stop AI data exfiltration.
• The engine achieved a zero percent attack success rate on Bench-Corp and AgentThreatBench.
• OpenAPPA runs outside the agent loop using deterministic rules and explicit recovery semantics.
Read more
OpenAPPA provides a deterministic approach to agent security that avoids the limitations of probabilistic judges while maintaining task completion utility.
• Archestra released the open-source OpenAPPA security engine to stop AI data exfiltration.
• The engine achieved a zero percent attack success rate on Bench-Corp and AgentThreatBench.
• OpenAPPA runs outside the agent loop using deterministic rules and explicit recovery semantics.
Read more
Google Restricts Gemini Models Across Free and Cheaper Tiers Starting October 2026
The shift changes what capabilities users get at different price points, bringing Google closer to the subscription structures used by competitors like OpenAI.
• Google will restrict free personal users to the Flash-Lite model starting in October 2026.
• Subscribers paying $4.99 per month for AI Plus will lose access to the Pro model.
• Access to all three models will require the AI Pro or AI Ultra subscription tiers.
Read more
The shift changes what capabilities users get at different price points, bringing Google closer to the subscription structures used by competitors like OpenAI.
• Google will restrict free personal users to the Flash-Lite model starting in October 2026.
• Subscribers paying $4.99 per month for AI Plus will lose access to the Pro model.
• Access to all three models will require the AI Pro or AI Ultra subscription tiers.
Read more
NASA and IBM Release Open Source Lunar Foundation Model
The model makes decades of complex lunar observation data usable for machine learning, improving tasks like polar ice deposit prediction and crater detection for future research.
• NASA and IBM released an open-source lunar foundation model using 17 years of orbiter data to aid machine learning.
• The model reduced prediction error for polar ice deposits by up to 22 percent compared to the best baseline, according to IBM.
• The model is publicly available on Hugging Face, with code hosted on GitHub and integrated into TerraTorch.
Read more
The model makes decades of complex lunar observation data usable for machine learning, improving tasks like polar ice deposit prediction and crater detection for future research.
• NASA and IBM released an open-source lunar foundation model using 17 years of orbiter data to aid machine learning.
• The model reduced prediction error for polar ice deposits by up to 22 percent compared to the best baseline, according to IBM.
• The model is publicly available on Hugging Face, with code hosted on GitHub and integrated into TerraTorch.
Read more
Chinese AI Models Echo State Doctrine on Sensitive Topics
The study highlights how training data and government regulations cause language models to reflect specific political ideologies, which could influence billions of users.
• A study by Aleph Alpha reveals Chinese AI models parrot state doctrine or refuse to answer on sensitive topics.
• The benchmark tested 967 taboo topics and rated only 17 to 41 percent of responses as balanced.
• Similar ideological patterns appeared in an Nvidia model trained partly on data generated by Chinese systems.
Read more
The study highlights how training data and government regulations cause language models to reflect specific political ideologies, which could influence billions of users.
• A study by Aleph Alpha reveals Chinese AI models parrot state doctrine or refuse to answer on sensitive topics.
• The benchmark tested 967 taboo topics and rated only 17 to 41 percent of responses as balanced.
• Similar ideological patterns appeared in an Nvidia model trained partly on data generated by Chinese systems.
Read more
Anthropic Asks Claude Users to Share Voice Data for AI Model Training
The feature gives users independent control over whether their voice conversations are used for artificial intelligence training separate from regular text chats.
• Anthropic now asks Claude users to voluntarily share voice data to help train and improve its artificial intelligence models.
• The voice training option is optional and disabled by default in the Privacy settings menu of the application.
• Users can delete their voice data from settings or turn the training feature off at any time.
Read more
The feature gives users independent control over whether their voice conversations are used for artificial intelligence training separate from regular text chats.
• Anthropic now asks Claude users to voluntarily share voice data to help train and improve its artificial intelligence models.
• The voice training option is optional and disabled by default in the Privacy settings menu of the application.
• Users can delete their voice data from settings or turn the training feature off at any time.
Read more
Trump Announces New Super Intelligence Force For AI Coordination
The initiative establishes a direct presidential reporting line for coordinating artificial intelligence policy and industry outreach across the government.
• President Donald Trump announced the creation of a Super Intelligence Force to coordinate artificial intelligence outreach.
• The new unit will manage communications with consumers, religious organizations, infrastructure operators, and AI companies.
• Intelligence director Jay Clayton and FTC chair Andrew Ferguson will lead the group reporting to the president.
Read more
The initiative establishes a direct presidential reporting line for coordinating artificial intelligence policy and industry outreach across the government.
• President Donald Trump announced the creation of a Super Intelligence Force to coordinate artificial intelligence outreach.
• The new unit will manage communications with consumers, religious organizations, infrastructure operators, and AI companies.
• Intelligence director Jay Clayton and FTC chair Andrew Ferguson will lead the group reporting to the president.
Read more
🤡1
Google Researchers Propose New Method to Prevent AI Agent Test Memorization
The method prevents AI agents from overspecializing on training data, helping them perform better on unseen tasks according to the study.
• Google researchers developed a method to stop self-improving AI agents from memorizing their tests.
• The approach uses a shrinking edit budget and a critic to prevent benchmark-specific memorization.
• Tests showed the method improved scores on unseen tasks while using fewer tokens.
Read more
The method prevents AI agents from overspecializing on training data, helping them perform better on unseen tasks according to the study.
• Google researchers developed a method to stop self-improving AI agents from memorizing their tests.
• The approach uses a shrinking edit budget and a critic to prevent benchmark-specific memorization.
• Tests showed the method improved scores on unseen tasks while using fewer tokens.
Read more
Ollama makes MLX the default for supported models on Apple Silicon
Apple Silicon users running architectures supported by MLX now get that runtime automatically, according to Ollama.
• Ollama says MLX is now the default for supported model architectures on Apple Silicon.
• The release notes also name gemma4, qwen3.6 and qwen3.5.
• Ollama says it will keep testing and enabling additional models.
Read more
Apple Silicon users running architectures supported by MLX now get that runtime automatically, according to Ollama.
• Ollama says MLX is now the default for supported model architectures on Apple Silicon.
• The release notes also name gemma4, qwen3.6 and qwen3.5.
• Ollama says it will keep testing and enabling additional models.
Read more
Google Pauses Open Source Bug Bounty Program Due to AI Submissions
The pause shows how AI generated noise can overwhelm security systems and force major tech companies to halt vulnerability reporting.
• Google paused its open source bug bounty program on October 1.
• The company cited a significant rise in invalid automated submissions and AI hallucinations.
• Google promised an update on the program status in the first quarter of 2027.
Read more
The pause shows how AI generated noise can overwhelm security systems and force major tech companies to halt vulnerability reporting.
• Google paused its open source bug bounty program on October 1.
• The company cited a significant rise in invalid automated submissions and AI hallucinations.
• Google promised an update on the program status in the first quarter of 2027.
Read more
OpenAI Introduces Visual Ads to ChatGPT During Image Generation
The shift lets OpenAI monetize its 1.2 billion weekly users through visual advertisements while implementing safeguards and conversion tracking with independent partners.
• OpenAI will test visual ads in ChatGPT during image generation later this month in the U.S.
• The ads will be clearly labeled, kept separate from generated images, and will not influence chatbot answers.
• OpenAI is partnering with analytics firms for conversion tracking and brand safety while protecting private user chats.
Read more
The shift lets OpenAI monetize its 1.2 billion weekly users through visual advertisements while implementing safeguards and conversion tracking with independent partners.
• OpenAI will test visual ads in ChatGPT during image generation later this month in the U.S.
• The ads will be clearly labeled, kept separate from generated images, and will not influence chatbot answers.
• OpenAI is partnering with analytics firms for conversion tracking and brand safety while protecting private user chats.
Read more