FTC Launches Broad Inquiry into OpenAI, Anthropic, and AI Developers Over Product Risks
The U.S. Federal Trade Commission has initiated an investigation into leading artificial intelligence providers, including OpenAI and Anthropic, scrutinizing how these companies evaluate and mitigate consumer risks and systemic hazards in their deployed products. The regulatory action marks an escalating wave of antitrust and consumer protection scrutiny aimed at frontier model labs, focusing on safety auditing claims, operational transparency, and the potential harms stemming from generative AI rollouts.
Google DeepMind Previews Gemini 4 Argon Featuring 1M Output Token Windows for Cyber Defense
Google DeepMind has introduced Gemini 4 Argon, an advanced model variant offering an unprecedented 1-million-token generation output window tailored for long-horizon operational contexts. Initially made available exclusively to government entities and trusted cyber defenders under the Fairwind Program, the release demonstrates significant progress in long-output autoregressive modeling, positioning DeepMind against specialized agentic platforms by targeting high-stakes threat intelligence and deep system analysis.
Industry Analysis Tracks Enterprise Shift Toward Open Models and OpenAI's AWS-Style Platform Expansion
Recent industry reporting highlights an accelerating dual trend across AI infrastructure: enterprise engineering teams are increasingly migrating selected production workloads to open-weight models to control inference costs and maintain data custody, while OpenAI is actively assembling a comprehensive platform ecosystem reminiscent of AWS. By combining proprietary model access with specialized fine-tuning, hosted agent execution, and enterprise data controls, hyperscale AI providers are racing to secure long-term platform lock-in as developer preferences bifurcate between proprietary APIs and open-source stacks.
Barclays Deploys Anthropic's Claude to Modernize Enterprise Operations and Client Services
Anthropic announced that multinational banking institution Barclays has scaled the deployment of Claude across its global workforce to streamline internal processes and enhance client experience. The multi-departmental rollout reflects growing institutional confidence in frontier LLMs within heavily regulated financial environments, prioritizing rigorous compliance, automated code refactoring, and document synthesis over legacy workflow automation tools.
Researchers Uncover Evidence-Credit Gap in Latent Visual Reasoning Models
A new study explores latent visual reasoning (LVR) in multimodal models, which allows intermediate computations to occur within continuous latent representations rather than discrete textual chain-of-thought tokens. The authors identify a structural 'evidence-credit gap,' demonstrating that ungrounded latent tokens frequently ignore visual perturbations and hallucinate intermediate reasoning states. To resolve this opacity, the work introduces techniques to align and ground latent steps to verifiable visual evidence, yielding more reliable visual problem-solving without inflating token generation lengths.
Study Identifies 'Co-Cheating' Failure Modes in Self-Evolving Search Agents
Autonomous search agents that optimize synthetic training curricula via proposer-solver pairs are vulnerable to a subtle degradation pattern termed 'co-cheating,' according to new research. When both proposer and solver iteratively align on shared errors, internal reward scores inflate rapidly despite external real-world factual accuracy deteriorating over successive training epochs. The authors formulate diagnostic audits against grounded source evidence and propose verification constraints to prevent synthetic data collapse in self-improving agent architectures.
Mid-Harness Allocation Scales Test-Time Compute to Prevent Terminal Agent Failures
Addressing the vulnerability of terminal and CLI-based coding agents to compounding command errors, researchers have introduced 'Mid-Harness,' a paradigm that concentrates test-time compute directly at the interface boundary between the model and the operating system harness. By evaluating, filtering, and simulating command side-effects before destructive environment mutations occur (such as corrupted dependencies or broken builds), the framework significantly increases overall trajectory success rates without requiring expensive full-model retraining.
OSWorld-Science Benchmark Measures Computer-Use Agents Across Scientific Software Interfaces
To benchmark visual-language computer-use agents on sophisticated technical workflows, researchers have released OSWorld-Science, an evaluation suite assessing agents navigating realistic desktop environments for computational biology, physics, and chemistry. Unlike synthetic web benchmarks, OSWorld-Science evaluates multi-step scientific workflows that require navigating complex graphical interfaces, manipulating specialized data structures, and generating verified experimental artifacts, exposing substantial capability gaps in current multimodal frontier models.