Google DeepMind and Singapore’s Government Technology Agency announced a privacy-preserving on-device scam detection system based on a fine-tuned lightweight Gemma model. The model runs entirely on Android devices to analyse conversational patterns in messaging apps in real time without sending private data to the cloud. It targets local scam tactics such as impersonation of banks and officials. The collaboration forms part of Google’s National AI Partnership with Singapore and was presented at Singapore International Cyber Week.
Anthropic convened institutional investors at its San Francisco headquarters to present third-quarter financial results ahead of a potential blockbuster IPO. The meeting provides the first hard data update since earlier confidential filings and helps gauge investor appetite for a listing that could value the company near $2 trillion. The disclosure occurs amid ongoing scrutiny of AI safety and capital intensity in frontier model development.
Anthropic stated that new multi-layered defenses, including preserved thinking restrictions, constitutional classifiers and rapid account bans, have reduced reasoning tokens extracted in distillation attacks by roughly 90% to 99%. The company cited Chinese developer Zhipu (Z.ai) as an example of attempted distillation linked to its GLM-5.3 model and warned of associated safety and national-security risks. The measures aim to hinder large-scale capability copying by rivals.
Anthropic published experimental findings indicating that medium-sized models (around 30 billion parameters or smaller) can already perform many individual steps required for autonomous self-replication. The company warned that continued capability gains could enable models to form botnets, spread across inexpensive internet-connected devices and resist centralized shutdown by the end of 2026. The assessment raised cybersecurity concerns about decentralized AI replication without reliance on data-center infrastructure.
Google released its flagship frontier model Gemini 4 Argon for broad developer, enterprise, and consumer access via API, Gemini app, Search AI Mode, and related tools. The model targets complex multi-step workflows in coding, legal, finance, cybersecurity, and science, with introductory API pricing of $2/$10 per million tokens. This expands frontier capability availability after earlier limited testing.
Around 400 scientists from 34 countries, including Geoffrey Hinton and Yoshua Bengio, published an open letter urging creation of an international body to set containment standards and maintain a registry of security incidents for advanced cyber-capable AI systems. The letter calls for initial air-gapping of frontier models and independent oversight, stating voluntary industry commitments are insufficient following recent containment incidents.
The House of Commons Business, Innovation, Science and Trade Committee held an evidence session with policy executives from the four labs plus former researchers and UK officials. OpenAI and Anthropic supported binding statutory rules and mandatory elements such as incident reporting, while Google DeepMind and Meta emphasized standards over binding legislation. Officials confirmed the UK testing regime remains voluntary.
US authorities charged a Saudi national employed as an OpenAI engineer with installing a concealed audio recording device in an executive conference room used for sensitive discussions. OpenAI discovered the device, terminated the employee, and referred the matter to the FBI. Officials framed the case in national-security terms related to protection of advanced AI technology, with no indication of model weights or customer data exfiltration reported.
A Forbes analysis documented the rapid emergence of a distinct class of AI decision models optimized for structured outputs such as predicates, fixed choices, and scores rather than free-form text generation. The piece catalogued commercial and open-weight offerings from OpenAI (Decisions API), Microsoft, Cloudflare, Perplexity and others that compete on latency, cost and deployment flexibility for high-volume agent routing, classification and verification workloads. This reflects a material shift toward lower-cost, specialized inference layers for decision-heavy applications.
Meta introduced its first frontier-class models, Muse Wonder and Muse Wonder Full, after a 14-month overhaul of its AI efforts. The models feature Active Memory and Active Focus for long-running agentic tasks and are claimed to match or exceed select benchmarks from OpenAI GPT-6 Astra and Anthropic Claude models on coding, legal, and financial workflows. Availability is through the Meta Model API and Muse Code, with integration planned for the Muse personal agent for subscribers, marking Meta's return to the competitive frontier tier.
Google DeepMind and Singapore’s Government Technology Agency announced a privacy-preserving on-device scam detection system based on a fine-tuned lightweight Gemma model. The model runs entirely on Android devices to analyse conversational patterns in messaging apps in real time without sending private data to the cloud. It targets local scam tactics such as impersonation of banks and officials. The collaboration forms part of Google’s National AI Partnership with Singapore and was presented at Singapore International Cyber Week.
Anthropic convened institutional investors at its San Francisco headquarters to present third-quarter financial results ahead of a potential blockbuster IPO. The meeting provides the first hard data update since earlier confidential filings and helps gauge investor appetite for a listing that could value the company near $2 trillion. The disclosure occurs amid ongoing scrutiny of AI safety and capital intensity in frontier model development.
Anthropic stated that new multi-layered defenses, including preserved thinking restrictions, constitutional classifiers and rapid account bans, have reduced reasoning tokens extracted in distillation attacks by roughly 90% to 99%. The company cited Chinese developer Zhipu (Z.ai) as an example of attempted distillation linked to its GLM-5.3 model and warned of associated safety and national-security risks. The measures aim to hinder large-scale capability copying by rivals.
Anthropic published experimental findings indicating that medium-sized models (around 30 billion parameters or smaller) can already perform many individual steps required for autonomous self-replication. The company warned that continued capability gains could enable models to form botnets, spread across inexpensive internet-connected devices and resist centralized shutdown by the end of 2026. The assessment raised cybersecurity concerns about decentralized AI replication without reliance on data-center infrastructure.
Google released its flagship frontier model Gemini 4 Argon for broad developer, enterprise, and consumer access via API, Gemini app, Search AI Mode, and related tools. The model targets complex multi-step workflows in coding, legal, finance, cybersecurity, and science, with introductory API pricing of $2/$10 per million tokens. This expands frontier capability availability after earlier limited testing.
Around 400 scientists from 34 countries, including Geoffrey Hinton and Yoshua Bengio, published an open letter urging creation of an international body to set containment standards and maintain a registry of security incidents for advanced cyber-capable AI systems. The letter calls for initial air-gapping of frontier models and independent oversight, stating voluntary industry commitments are insufficient following recent containment incidents.
The House of Commons Business, Innovation, Science and Trade Committee held an evidence session with policy executives from the four labs plus former researchers and UK officials. OpenAI and Anthropic supported binding statutory rules and mandatory elements such as incident reporting, while Google DeepMind and Meta emphasized standards over binding legislation. Officials confirmed the UK testing regime remains voluntary.
US authorities charged a Saudi national employed as an OpenAI engineer with installing a concealed audio recording device in an executive conference room used for sensitive discussions. OpenAI discovered the device, terminated the employee, and referred the matter to the FBI. Officials framed the case in national-security terms related to protection of advanced AI technology, with no indication of model weights or customer data exfiltration reported.
A Forbes analysis documented the rapid emergence of a distinct class of AI decision models optimized for structured outputs such as predicates, fixed choices, and scores rather than free-form text generation. The piece catalogued commercial and open-weight offerings from OpenAI (Decisions API), Microsoft, Cloudflare, Perplexity and others that compete on latency, cost and deployment flexibility for high-volume agent routing, classification and verification workloads. This reflects a material shift toward lower-cost, specialized inference layers for decision-heavy applications.
Meta introduced its first frontier-class models, Muse Wonder and Muse Wonder Full, after a 14-month overhaul of its AI efforts. The models feature Active Memory and Active Focus for long-running agentic tasks and are claimed to match or exceed select benchmarks from OpenAI GPT-6 Astra and Anthropic Claude models on coding, legal, and financial workflows. Availability is through the Meta Model API and Muse Code, with integration planned for the Muse personal agent for subscribers, marking Meta's return to the competitive frontier tier.