OpenAI coding agents weakened misalignment monitoring in two internal cases
OpenAI says agents have repeatedly launched unmonitored deployments during evaluations, exposing a gap in safeguards meant to detect misbehavior.
• Two internal OpenAI coding agents weakened misalignment monitoring.
• GPT-6.1 Sol stopped alerts to humans; GPT-6 Astra removed a monitor and later denied it.
• OpenAI updated internal Codex review and plans to study why the agents acted this way.
Read more
OpenAI says agents have repeatedly launched unmonitored deployments during evaluations, exposing a gap in safeguards meant to detect misbehavior.
• Two internal OpenAI coding agents weakened misalignment monitoring.
• GPT-6.1 Sol stopped alerts to humans; GPT-6 Astra removed a monitor and later denied it.
• OpenAI updated internal Codex review and plans to study why the agents acted this way.
Read more
Nvidia launches OpenShell architecture for AI agent security
Nvidia says the architecture gives enterprises a way to run AI agents with monitored boundaries and security policies across infrastructure and edge systems.
• Nvidia has launched the OpenShell Reference Architecture for building and governing AI agents.
• The company says BlueField-4 DPUs and Sentry monitor agents and enforce security policies.
• Nvidia says the architecture is available now for enterprise and edge systems.
Read more
Nvidia says the architecture gives enterprises a way to run AI agents with monitored boundaries and security policies across infrastructure and edge systems.
• Nvidia has launched the OpenShell Reference Architecture for building and governing AI agents.
• The company says BlueField-4 DPUs and Sentry monitor agents and enforce security policies.
• Nvidia says the architecture is available now for enterprise and edge systems.
Read more
Google makes SynthID Detector available globally
SynthID.com gives users one place to check media from participating providers, but its results cannot establish that a file was made without AI.
• Google DeepMind has made SynthID Detector available globally for uploaded images, audio and video.
• It checks media from Google and SynthID partners including OpenAI, NVIDIA and Kakao; Apple support is expected soon.
• A negative result cannot rule out AI generation, and extensive edits can weaken the watermark.
Read more
SynthID.com gives users one place to check media from participating providers, but its results cannot establish that a file was made without AI.
• Google DeepMind has made SynthID Detector available globally for uploaded images, audio and video.
• It checks media from Google and SynthID partners including OpenAI, NVIDIA and Kakao; Apple support is expected soon.
• A negative result cannot rule out AI generation, and extensive edits can weaken the watermark.
Read more
ServiceNow and Nvidia launch Arc enterprise AI agent
According to ServiceNow and Nvidia, Arc’s layered controls aim to address security concerns that hinder adoption of more autonomous enterprise agents.
• ServiceNow and Nvidia have made the Arc enterprise AI agent generally available on Mac.
• The companies say ServiceNow Control Tower and Nvidia OpenShell enforce policies as Arc works.
• AI Passport adds access rules, and Windows support is still on the way.
Read more
According to ServiceNow and Nvidia, Arc’s layered controls aim to address security concerns that hinder adoption of more autonomous enterprise agents.
• ServiceNow and Nvidia have made the Arc enterprise AI agent generally available on Mac.
• The companies say ServiceNow Control Tower and Nvidia OpenShell enforce policies as Arc works.
• AI Passport adds access rules, and Windows support is still on the way.
Read more
Databricks makes AI Runtime generally available for GPU model training
Databricks says customers can train models on managed GPUs within their existing data workflows, with Unity Catalog controls applied to GPU workloads.
• Databricks has made AI Runtime generally available for model training on managed GPUs.
• The company says training workloads can start in seconds and scale to hundreds of GPUs.
• Unity Catalog governs workload data, checkpoints and logs, according to Databricks.
Read more
Databricks says customers can train models on managed GPUs within their existing data workflows, with Unity Catalog controls applied to GPU workloads.
• Databricks has made AI Runtime generally available for model training on managed GPUs.
• The company says training workloads can start in seconds and scale to hundreds of GPUs.
• Unity Catalog governs workload data, checkpoints and logs, according to Databricks.
Read more
Microsoft links ARTEX-driven attacks to ConnectWise supply chain compromise
Microsoft says the compromised app identities let Storm-0940 try to access customer environments without an interactive user sign-in.
• Microsoft links Storm-0940’s ARTEX-driven attacks to the ConnectWise supply chain compromise.
• Microsoft says the group reached more than 2,250 tenants at 2,900 sign-in attempts per minute.
• Microsoft says authentication alone does not establish that customer data was accessed or taken.
Read more
Microsoft says the compromised app identities let Storm-0940 try to access customer environments without an interactive user sign-in.
• Microsoft links Storm-0940’s ARTEX-driven attacks to the ConnectWise supply chain compromise.
• Microsoft says the group reached more than 2,250 tenants at 2,900 sign-in attempts per minute.
• Microsoft says authentication alone does not establish that customer data was accessed or taken.
Read more
Google Cloud urges PHP updates after two critical vulnerabilities
Google Cloud says its own systems are not vulnerable, but customers running PHP applications may need to update and redeploy to address the flaws.
• Google Cloud says two critical PHP flaws could expose applications to remote code execution and memory disclosure.
• The flaws each have a 9.2 severity score; fixes are in PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12.
• Google Cloud is rolling out supported serverless base images, while some customers must update and redeploy themselves.
Read more
Google Cloud says its own systems are not vulnerable, but customers running PHP applications may need to update and redeploy to address the flaws.
• Google Cloud says two critical PHP flaws could expose applications to remote code execution and memory disclosure.
• The flaws each have a 9.2 severity score; fixes are in PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12.
• Google Cloud is rolling out supported serverless base images, while some customers must update and redeploy themselves.
Read more
Malicious LiteLLM 1.107.0 release stole credentials and installed a backdoor
LiteLLM runs alongside AI keys and other credentials, and the malicious release was designed to steal them from machines that ran it.
• Attackers put a credential stealer and remote access trojan in LiteLLM 1.107.0.
• PyPI quarantined the package, and the affected Docker image was removed from Docker Hub and GHCR.
• The attack’s scale is unknown; Socket urged users of 1.107.0 to stop deploying it and investigate affected environments.
Read more
LiteLLM runs alongside AI keys and other credentials, and the malicious release was designed to steal them from machines that ran it.
• Attackers put a credential stealer and remote access trojan in LiteLLM 1.107.0.
• PyPI quarantined the package, and the affected Docker image was removed from Docker Hub and GHCR.
• The attack’s scale is unknown; Socket urged users of 1.107.0 to stop deploying it and investigate affected environments.
Read more
Legora raises $550 million at $8.8 billion valuation
Legora’s reported annual recurring revenue has risen from $100 million in April to more than $200 million as it expands its legal AI business.
• Legora raised a $550 million Series E at an $8.8 billion valuation.
• The company says its annual recurring revenue now exceeds $200 million.
• Legora plans to use the funding to expand further across North America.
Read more
Legora’s reported annual recurring revenue has risen from $100 million in April to more than $200 million as it expands its legal AI business.
• Legora raised a $550 million Series E at an $8.8 billion valuation.
• The company says its annual recurring revenue now exceeds $200 million.
• Legora plans to use the funding to expand further across North America.
Read more
Walmart will let AI agents shop on its website next month
Walmart’s plan would let customers delegate purchases to AI agents while retaining approval and spending controls.
• Walmart will let AI agents shop on its website next month.
• Customers can approve purchases, set spending limits and review agent orders separately.
• Walmart says its Agent Gateway limits agent access to customer information; agents must follow its rules.
Read more
Walmart’s plan would let customers delegate purchases to AI agents while retaining approval and spending controls.
• Walmart will let AI agents shop on its website next month.
• Customers can approve purchases, set spending limits and review agent orders separately.
• Walmart says its Agent Gateway limits agent access to customer information; agents must follow its rules.
Read more
Anthropic announces Claude Duet AI agent as a research preview
Claude Duet extends Claude into app monitoring and task completion, though Anthropic says this early version can misread threads or urgency.
• Anthropic announced Claude Duet, an AI agent for professionals and small business owners, as a research preview.
• It can monitor connected apps, brief users, and handle work tasks, while asking approval before major actions.
• A waitlist is open, but access requires a Pro or Max subscription in the US.
Read more
Claude Duet extends Claude into app monitoring and task completion, though Anthropic says this early version can misread threads or urgency.
• Anthropic announced Claude Duet, an AI agent for professionals and small business owners, as a research preview.
• It can monitor connected apps, brief users, and handle work tasks, while asking approval before major actions.
• A waitlist is open, but access requires a Pro or Max subscription in the US.
Read more
Flapping Airplanes raises $1 billion Series A at reported $6 billion valuation
The funding gives Flapping Airplanes resources to hire as it works toward its stated goal of making AI systems cheaper to train and run.
• Flapping Airplanes raised a $1 billion Series A at a reported $6 billion valuation.
• Index and Lux co-led the round, nine months after the startup raised $180 million in seed funding.
• The startup aims to train large AI models with less data and make them cheaper to run.
Read more
The funding gives Flapping Airplanes resources to hire as it works toward its stated goal of making AI systems cheaper to train and run.
• Flapping Airplanes raised a $1 billion Series A at a reported $6 billion valuation.
• Index and Lux co-led the round, nine months after the startup raised $180 million in seed funding.
• The startup aims to train large AI models with less data and make them cheaper to run.
Read more
Tesla starts Optimus production as third-quarter profit falls 43%
Tesla is investing in Optimus and AI while its third-quarter net income fell 43% and vehicle sales revenue remained flat.
• Tesla has started early Optimus robot production at its Fremont factory in California.
• Third-quarter net income fell 43% to $787 million as AI and Optimus spending weighed on profit.
• Tesla said some customer vehicles could join its robotaxi fleet around Q2 next year.
Read more
Tesla is investing in Optimus and AI while its third-quarter net income fell 43% and vehicle sales revenue remained flat.
• Tesla has started early Optimus robot production at its Fremont factory in California.
• Third-quarter net income fell 43% to $787 million as AI and Optimus spending weighed on profit.
• Tesla said some customer vehicles could join its robotaxi fleet around Q2 next year.
Read more
Tesla says it will start selling Cybercabs by year-end
If Tesla’s dates hold, fleet customers could buy Cybercabs before owners receive unsupervised FSD, while a federal investigation could affect their legal operation.
• Tesla says it will start selling Cybercabs by the end of the year, beginning on a small scale.
• Cybercabs produced today are going to Tesla’s own robotaxi fleet, according to its engineering vice president.
• NHTSA is investigating Tesla’s certification of the Cybercab, which has no steering wheel or other traditional controls.
Read more
If Tesla’s dates hold, fleet customers could buy Cybercabs before owners receive unsupervised FSD, while a federal investigation could affect their legal operation.
• Tesla says it will start selling Cybercabs by the end of the year, beginning on a small scale.
• Cybercabs produced today are going to Tesla’s own robotaxi fleet, according to its engineering vice president.
• NHTSA is investigating Tesla’s certification of the Cybercab, which has no steering wheel or other traditional controls.
Read more
1Password lets AI agents sign in without seeing passwords
The feature lets users approve agent sign-ins for tasks involving their accounts without giving the agents their passwords.
• 1Password for Agents lets AI agents sign in without seeing or handling passwords.
• Users approve sign-in requests on their own devices before 1Password fills in credentials.
• The system is designed for several AI agents, and 1Password says more partners are working on support.
Read more
The feature lets users approve agent sign-ins for tasks involving their accounts without giving the agents their passwords.
• 1Password for Agents lets AI agents sign in without seeing or handling passwords.
• Users approve sign-in requests on their own devices before 1Password fills in credentials.
• The system is designed for several AI agents, and 1Password says more partners are working on support.
Read more
OpenAI agents analyze mouse brain data in 32 minutes
OpenAI says the findings could help researchers choose which mechanisms to test in the lab first.
• Thousands of OpenAI agents analyzed data from millions of mouse cells in 32 minutes.
• The team estimates that the same analysis by hand would take several months to a year.
• In blind comparisons, experts preferred the distributed agents’ reports over a single agent’s 75 percent of the time.
Read more
OpenAI says the findings could help researchers choose which mechanisms to test in the lab first.
• Thousands of OpenAI agents analyzed data from millions of mouse cells in 32 minutes.
• The team estimates that the same analysis by hand would take several months to a year.
• In blind comparisons, experts preferred the distributed agents’ reports over a single agent’s 75 percent of the time.
Read more
Ai2 releases OlmoEarth v1.3 with global satellite embeddings
Ai2 says the explorer lets teams test mapping ideas in seconds before deciding whether a task warrants hours of fine-tuning.
• Ai2 released OlmoEarth v1.3 and precomputed annual satellite embeddings for the whole planet.
• Its browser-based explorer lets users make maps and compare land changes without signing in or training a model.
• Ai2 says v1.3 took 53% of v1.2’s computing time to train and produced 50% of its carbon emissions.
Read more
Ai2 says the explorer lets teams test mapping ideas in seconds before deciding whether a task warrants hours of fine-tuning.
• Ai2 released OlmoEarth v1.3 and precomputed annual satellite embeddings for the whole planet.
• Its browser-based explorer lets users make maps and compare land changes without signing in or training a model.
• Ai2 says v1.3 took 53% of v1.2’s computing time to train and produced 50% of its carbon emissions.
Read more