The Informed Layer · Newsscan
Daily briefing

Agent security risks rise; Regulators eye liability

← All issues

Anthropic / Claude ecosystem

Claude Computes N=4 Super Yang-Mills Amplitude to Nine Loops in Physics Breakthrough

Anthropic published research demonstrating that Claude successfully computed the N=4 super Yang-Mills scattering amplitude to nine loops. The milestone resolves an open computational physics challenge that had remained intractable under classical analytic and numerical techniques.

Anthropic Releases Claude Code v2.1.281 with Guardrail and IAM Role Enhancements

Anthropic shipped Claude Code v2.1.281, introducing 176 changes including enhanced confirmation safeguards against destructive shell commands and support for IAM role assumption in Bedrock gateways. The update also adds a dedicated boolean attribution flag to simplify provenance management in enterprise codebases.

Anthropic Patches Claude Code Bug Silently Disabling Local AGENTS.md Rules

Anthropic resolved a vulnerability in Claude Code where opting out of telemetry silently prevented the agent harness from loading local AGENTS.md instruction files. The issue previously risked bypassing developer-defined repository constraints and security rules during local agent execution.

Frontier model providers

OpenAI to Preview GPT-6 Cyber Model and Enterprise Security Suite at DevDay

OpenAI is preparing to preview GPT-6 Cyber alongside a dedicated deployment platform designed to automate enterprise vulnerability detection and remediation. The specialized model is targeted at defending systems against automated AI cyberattacks and will be integrated into restricted enterprise defensive programs.

AI developer tooling & infrastructure

Cohere Launches Managed Compass Cloud Beta for Enterprise Retrieval

Cohere introduced Compass Cloud into private beta, offering its enterprise retrieval engine as a fully managed SaaS service alongside its existing self-hosted offering. The service is tailored for multi-hop agentic search and complex enterprise context retrieval while minimizing self-hosting overhead.

Anysphere Ships Cursor 1.4 with Multi-Model Agent Support and Audit APIs

Anysphere released Cursor 1.4, adding granular multi-agent execution controls, multi-model tool routing, and deep GitHub pull request automation. The update also includes an enterprise AI code-tracking API to monitor and audit agent-generated source code across repositories.

GitGuardian Uncovers 24,000+ Secrets Leaked Across AI Coding Agent Endpoints

A security investigation by GitGuardian revealed that AI coding assistants including Cursor, Claude Code, and GitHub Copilot frequently expose API keys and credentials across unscanned local configuration files and shell logs. The study identified over 24,000 unique secrets in public MCP configurations, with more than 2,100 verified as active.

LangChain Unveils LangSmith Engine v2 with Automated Red Teaming

LangChain launched LangSmith Engine v2, adding automated red teaming, proactive vulnerability detection, and performance benchmarking for production agents. The platform automatically identifies failure points in agentic reasoning loops and generates validated fix proposals prior to production deployment.

Model Context Protocol Specification Updated for Stateless Serverless Deployment

The Model Context Protocol project and AWS announced an update removing protocol-level session affinity requirements from MCP server specifications. The stateless architecture allows developers to host MCP endpoints on AWS Lambda and serverless container runtimes without maintaining persistent sticky connections.

Archipelo Launches Salmon Cryptographic Execution Verification for AI Agents

Security firm Archipelo introduced Salmon, an Execution Verification Infrastructure (EVI) designed to provide cryptographic proof of autonomous agent activity. The system creates immutable audit trails to prevent AI agents from tampering with logs or evading sandbox constraints during complex workflows.

Cloud & platform providers

Microsoft Revamps Copilot into Unified Enterprise Super-App with Agent Tools

Microsoft unveiled a major overhaul of its Copilot ecosystem, consolidating Chat, Code authoring, and autonomous Autopilot workflows into a unified workspace. The platform incorporates Fabric IQ and Work IQ context engines along with usage-based controls to manage enterprise AI sprawl and compute expenditures.

AI policy, regulation & governance

Open-Source AI Agent Swarm Breaches 27 Companies at $25 Per Target

Security researchers documented a cyber campaign where an automated toolchain of open-source AI agents (Strix, Cairn, and Hermes) breached 27 corporate networks and exfiltrated 600,000 payment cards. The autonomous attacks operated at an estimated infrastructure cost of just $25 per compromised target.

Federal Trade Commission leadership indicated that regulatory enforcement will treat AI developers as legally responsible for the autonomous conduct of their AI agents, rejecting legal arguments that agentic autonomy shields parent companies from consumer protection and liability statutes.

New York City Council Unveils Sweeping Municipal AI Safety Legislation

The New York City Council introduced a comprehensive legislative package governing frontier AI and agent safety. The proposed bills mandate third-party algorithmic validation, private rights of action for AI harms, whistleblower protection frameworks, and required kill-switch mechanisms for autonomous software.

US Senators Introduce AI Risk Management and Security Act Requiring 45-Day Pre-Release Audits

U.S. Senators Mark Warner, Brian Schatz, and Andy Kim introduced the Artificial Intelligence Risk Management and Security Act of 2026. The bill would mandate that frontier AI developers submit advanced models to federal authorities for security testing 45 days prior to release and establishes an AI Safety Board with fines up to $250,000 per day.

US Senator Markey Proposes Independent Board to Investigate AI Cyberattacks

Senator Edward Markey introduced the Cybersecurity and AI Board of Investigations Act to create an independent investigative body with subpoena powers dedicated to probing AI-driven cyber incidents and critical infrastructure breaches.

Australian Government Strengthens AI Guardrails Following Health System Agent Breach

Australian Prime Minister Anthony Albanese reinforced commitments to enforce mandatory national AI guardrails after an autonomous AI agent breach impacted public health systems. The incident has intensified bipartisan support for statutory duty-of-care legislation and stricter operational standards.

Industry & market moves

Anthropic Signs $11.6 Billion Multi-Year Compute Deal with Akamai

Anthropic has entered into an $11.6 billion, seven-year computing power agreement with Akamai, expanding a previous engagement and granting an equity warrant for up to a 5% stake. The deal represents Akamai's largest commercial contract to date and secures dedicated distributed infrastructure for Anthropic's growing frontier model training and inference workloads.

SoftBank Forms Joint Venture and Commits $225M to Autonomous Heavy Equipment Developer ASI

SoftBank Group formed a joint venture with Autonomous Solutions, Inc. (ASI) backed by a $225 million investment to deploy autonomous heavy equipment across global construction and infrastructure sites. The venture targets mixed-fleet operational automation.

Stakk Acquires Document AI Specialist ParaScript for $63 Million

Australian enterprise identity firm Stakk Limited acquired document processing company ParaScript for $63 million. The combined entity will integrate automated document classification with real-time behavioral fraud analysis to mitigate AI-generated financial fraud.

Huspy Expands into Italy with $86M Investment and Integra Finance Acquisition

UAE-based proptech firm Huspy announced an $86 million investment and the acquisition of Italian credit intermediary Integra Finance. The expansion aims to introduce AI-native mortgage processing and real estate agent workflow automation across Southern Europe.

AI product & feature launches

Adobe Expands Integration Suite Across Gemini and Claude Platforms

Adobe rolled out expanded integrations across Google Gemini and Anthropic Claude, embedding over 80 creative and document management tools. The update provides native connected app functionality for Acrobat, Express, and interactive creative workflows directly inside conversational interfaces.

Meta Enhances Muse Guardrails Following Account Exposure Vulnerability

Meta has upgraded security warnings and access controls for its personal AI agent Muse after researchers disclosed a vulnerability that allowed potential unauthorized access to virtualized cloud accounts. The issue was patched via Meta's bug bounty workflow before broad exploitation.

SentinelOne Extends Wayfinder AI Threat Hunting Across AWS, Azure, and GCP

SentinelOne expanded its Wayfinder AI-driven threat hunting service to monitor multi-cloud control planes across AWS, Microsoft Azure, and Google Cloud. The solution combines automated telemetry analysis with continuous agent-driven investigation of cloud identity and infrastructure compromise.

NVIDIA Introduces NV-Reason-CT for 3D Radiographic Reasoning

NVIDIA unveiled NV-Reason-CT, a vision-language reasoning model designed for full 3D computed tomography volumes. The model utilizes a dedicated 3D vision transformer to emulate radiologist diagnostic workflows, producing structured clinical reports and step-by-step diagnostic reasoning validated by clinical partners.

Black Forest Labs Unveils FLUX 3 Action Model for Humanoid Robotics

Black Forest Labs released FLUX 3 Action, a lightweight foundation model optimized for robotic control and manipulation. In benchmark evaluations on RoboLab-120, the model achieved a 42.92% success rate, outpacing comparable models while operating with 56% fewer parameters and 4x faster execution speeds.

Research with immediate practical relevance

AlphaFold Database Adds 2,800+ Viral Protein Complex Predictions for Pandemic Defense

Google DeepMind and the European Bioinformatics Institute (EMBL-EBI) have expanded the AlphaFold Database with structural predictions covering more than 2,800 viral protein complexes. The open-access dataset aims to accelerate drug and vaccine discovery for emerging infectious diseases and pandemic preparedness.

NVIDIA Open-Sources Nemotron 3 Diarization Model for Real-Time Speaker Transcription

NVIDIA released Nemotron 3 Diarization as an open-weight model capable of tracking up to eight speakers simultaneously in real-time streams. The model achieved first place on the Voice Arena benchmark with a 14.72% error rate, representing a ~24% error reduction over prior open baselines.

Tsinghua University Releases VeriLoop E2 27B Model with Full Quantization Ladder

Tsinghua SIGS Robot Lab open-sourced VeriLoop E2, a 27-billion parameter post-trained model featuring governed recurrence for verifiable code generation and reasoning. The release includes a complete GGUF precision ladder spanning BF16 to IQ1_M, delivering up to 66.5% memory footprint reduction with minimal perplexity degradation.