Anthropic / Claude ecosystem
Anthropic Warns Infostealer Malware is Hijacking Claude Login Sessions
Anthropic has disclosed that infostealer malware variants, including LummaC2, StealC, and Vidar, are actively harvesting browser session tokens to hijack authenticated Claude accounts. The company is forcibly revoking compromised sessions and refunding unauthorized credit consumption caused by illicit automated usage.
- Source: BleepingComputer
- Significance: Enterprises must enforce strict session lifetimes, device posture checks, and dedicated API key management to prevent malicious credential theft from draining organizational AI compute budgets.
Frontier model providers
OpenAI Deprecates Standalone DALL-E GPT Tool to Streamline Native Image Generation
OpenAI has officially shut down its dedicated DALL-E GPT tool as part of a portfolio streamlining initiative. Users have been instructed to download legacy outputs as image creation is consolidated entirely into native ChatGPT Images powered by newer generation models.
- Source: Inc.
- Significance: Confirms OpenAI's ongoing consolidation of standalone multimodal tools into unified chat interfaces, requiring teams using legacy custom GPT endpoints to migrate to modern native APIs.
Google DeepMind Demonstrates Autonomous Laboratory Experiment Execution with Gemini
Google DeepMind researchers published findings (arXiv:2608.26701) demonstrating that Gemini models can autonomously plan and execute physical laboratory experiments. By interfacing directly with robotic lab equipment, the system moves past purely computational hypothesis generation to conduct closed-loop wet-lab research.
- Source: 36Kr
- Significance: Expands the frontier of AI in life sciences and materials discovery by bridging digital reasoning models with automated physical execution hardware.
- Potentially previously reported: Google Deepmind's AI Co-Scientist now plans experiments, runs lab equipment, and writes scientific papers
Meta AI Introduces EvoHarness-RL to Match Frontier Model Capabilities on 8B Models
Meta AI researchers introduced EvoHarness-RL, a reinforcement learning framework that allows an 8-billion-parameter language model to match the benchmark performance of frontier models like Claude Opus 4.5. The approach uses targeted evolutionary search and reward modeling to drastically reduce the compute footprint required for complex reasoning.
- Source: It's The Coin
- Significance: Significantly lowers the hardware barrier for deploying high-accuracy domain-specific models locally or in cost-constrained enterprise private clouds.
- Potentially previously reported: Meta’s EvoHarness-RL Takes Qwen3-8B to 96.9% on ALFWorld, Near Claude Opus 4.5 | Superpower Daily
AI developer tooling & infrastructure
vLLM v0.28.0 Adds Shared-Expert Sharding and Memory Optimizations for Kimi-K3 and DeepSeek V4
Open-source LLM serving engine vLLM has released v0.28.0, featuring performance optimizations for Kimi-K3 and DeepSeek V4. The update introduces shared-expert sharding and sparse MLA optimizations, claiming a ~60% improvement in time-to-first-token (TTFT) and saving up to 17 GiB of VRAM per GPU.
- Source: AIToolly
- Significance: Drastically decreases the hardware cost and latency of serving next-generation Chinese open-weight frontier models in production environments.
Community Plugin Adds Inverted Model Selection Routing to DeepSeek Harness
Developers have released dsh-ui-mod, a third-party plugin for DeepSeek Harness that introduces inverted model selection and flexible fallback routing. The tool enables developers to configure granular failovers across local and remote endpoint clusters depending on model load and response latency.
- Source: PulseAugur
- Significance: Provides infrastructure teams using the DeepSeek ecosystem with customizable routing and load-balancing controls across multi-region model endpoints.
WebAZ Introduces Decoupled Contribution Architecture for Remote MCP AI Agents
WebAZ has launched a contribution management system for Remote Model Context Protocol (MCP) agents. The platform decouples automated agent task proposals from formal human review workflows, enabling organizations to safely intake AI-suggested code and data modifications without granting direct commit privileges.
- Source: PulseAugur
- Significance: Standardizes containment and review workflows for autonomous AI agents contributing changes across distributed enterprise repositories.
Cloud & platform providers
AWS Uncovers $100M+ Bedrock Credit Exploitation Scheme and Implements Emergency Controls
A major investigation revealed that rogue token brokers exploited AWS Bedrock's enterprise credit-override mechanism to siphon over $100 million in compute credits for illicit resale. In response, AWS has implemented emergency identity verification and dual-approval requirements across all promotional credit redemption paths.
- Source: CTOL Digital Solutions
- Significance: Cloud security teams must audit API token allocation pipelines and implement strict access monitoring to protect corporate cloud quotas from unauthorized provisioning.
AI policy, regulation & governance
OpenAI, Anthropic, and 100+ Organisations Sign Open Letter Urging Global Cyber Defence Overhaul
More than 100 technology organizations and frontier AI labs, including OpenAI and Anthropic, have co-signed an open letter warning that AI-enabled cyberattacks on critical infrastructure are approaching parity with advanced human threat actors. The coalition calls on industry and governments to mandate automated defenses, rapid patching protocols, and formal red-teaming disclosures.
- Source: Decrypt
- Significance: Signals a unified shift among frontier model developers toward advocating strict defensive cybersecurity standards as autonomous offensive agent capabilities rapidly mature.
- Potentially previously reported: OpenAI, Anthropic, Google, and 100 other companies call for action to defend against rogue AI | TechCrunch
Germany's BSI Finalises Consultation on AI Audit and Assurance Architecture (A5)
Germany's Federal Office for Information Security (BSI) has concluded its public consultation on the AI Audit and Assurance Assessment Architecture (A5). The framework establishes standardized technical criteria for evaluating the cybersecurity, transparency, and operational robustness of enterprise AI deployments.
- Source: Digital Policy Alert
- Significance: Enterprises operating in Germany and across the EU must prepare compliance and auditing pipelines against the BSI's emerging technical assessment architecture.
US Lawmakers Introduce Bipartisan Bill to Monitor Self-Improving Autonomous AI Systems
US Representative George Whitesides has introduced the bipartisan 'Self-Improving AI Monitoring Act', designed to give federal evaluators visibility into frontier AI labs utilizing autonomous self-improving models for recursive software engineering and model development. The legislation mandates structured reporting when AI systems independently generate new training or alignment procedures.
- Source: Hometown Station
- Significance: Frontier AI developers face prospective oversight and mandatory compliance disclosures regarding autonomous model improvement loops and recursive self-coding pipelines.
Australian Government to Overhaul Privacy Act Targeting AI Systems and Smart Glasses
The Australian Federal Government has announced comprehensive reforms to the Privacy Act specifically addressing biometric capture by AI systems and wearable smart glasses. The draft legislative package includes a statutory 'right to erasure' and heightened disclosure requirements for companies processing ambient video and audio data.
- Source: The Sydney Morning Herald
- Significance: Australian enterprises deploying AI vision systems or smart workplace wearables must adapt data governance frameworks to accommodate mandatory deletion requests and explicit user consent rules.
Australian Government Considers Mandatory Social Media Algorithm Opt-Outs
The Australian Government is reviewing proposals to introduce a digital 'duty of care' that would compel social media platforms to offer chronological, un-algorithmic feeds to users. The initiative aims to limit exposure to toxic or polarizing AI-curated recommendation loops.
- Source: The Guardian
- Significance: Consumer-facing digital platforms operating in Australia may need to re-engineer their core feed architectures to provide fully optional recommendation engines.
Australian eSafety Commissioner Warns of Doubling in Deepfake Complaints Across ACT Schools
The Australian eSafety Commissioner reported a 100% surge in complaints regarding AI-generated deepfake intimate imagery among school-aged students. The alert has prompted educational authorities in the Australian Capital Territory to recommend removing public student directory photographs to prevent synthetic misuse.
- Source: The Canberra Times
- Significance: Highlights rising regulatory pressure on generative media creation platforms to implement cryptographic watermarking and provenance safeguards.
Industry & market moves
USD.AI Secures $100M Stablecoin Debt Facility from Bullish to Fund GPU-Backed Lending
DeFi lending platform USD.AI (developed by Permian Labs) has secured a $100 million stablecoin debt facility from digital asset exchange Bullish. The facility will be used to originate loans collateralized by high-performance GPU hardware and enterprise AI data center compute contracts.
- Source: Techy101
- Significance: Reflects the emergence of compute-collateralized debt structures to bridge capital financing for neocloud operators and AI infrastructure builders.
- Potentially previously reported: USD.AI Secures $100M Stablecoin Debt Facility From Bullish for GPU Financing
IndiGo Ventures Backs Indian Sovereign AI Developer Sarvam in Series B Extension
IndiGo Ventures, the corporate venture arm of Indian carrier IndiGo, has made a strategic equity investment in Sarvam AI as part of Sarvam's Series B round. The partnership focuses on co-developing sovereign Indian language voice and automation agents for airline customer operations and maintenance workflows.
- Source: Aviation Defence Universe
- Significance: Demonstrates enterprise backing for localized, sovereign AI models in major non-English emerging markets.
- Potentially previously reported: Artificial Intelligence: IndiGo Ventures invests in Sarvam to enhance Artificial Intelligence in aviation - The Times of India
ITC Infotech Eyes ₹2,900 Crore Acquisition of Happiest Minds to Consolidate AI Services
ITC Infotech is reportedly preparing a ₹2,800–2,900 crore ($350M+) bid to acquire a 44% controlling stake in IT services firm Happiest Minds Technologies. The proposed transaction aims to expand ITC Infotech's generative AI, cloud migration, and automation delivery capabilities globally.
- Source: CIOL
- Significance: Signals continued consolidation across mid-tier IT service providers aiming to scale specialized AI implementation practices through M&A.
- Potentially previously reported: Happiest Minds shares fall 6% amid renewed ITC Infotech stake-sale buzz - The HinduBusinessLine
AI product & feature launches
Meta Unbundles Horizon App to Launch Dedicated 'Horizon Create' AI Tooling Suite
Meta is splitting its Horizon companion application into two distinct apps, spinning off mobile 3D world navigation and generative AI creation tools into a dedicated app titled 'Horizon Create'. The flagship Horizon app will pivot exclusively to VR device management and store discovery.
- Source: VR.org
- Significance: Highlights Meta's strategy to lower creator barriers by decoupling generative 3D world creation tools from dedicated VR headsets onto mobile devices.
Research with immediate practical relevance
MirroS Unveils 'Code-as-World' Framework Translating Real Video Into Executable MuJoCo Physics Programs
AI research group MirroS has released 'Code-as-World', an agentic system that reconstructs real-world video footage directly into executable MuJoCo physics code rather than raw pixel tokens. The system scored 55.4 MRA on the QuantiPhy benchmark, surpassing conventional vision-language models in physical dynamics prediction.
- Source: TradePoint
- Significance: Provides physical AI and simulation engineers with a method to convert real-world sensory capture into interactive, physically accurate digital twins for robot training.
NPR and NewsGuard Study Finds AI Chatbots Resist State Disinformation Better Than Search Engines
A comparative evaluation conducted by NPR and NewsGuard found that major conversational AI chatbots are significantly more effective at identifying and rejecting state-backed propaganda narratives than traditional search engines and algorithmic snippets.
- Source: GPB News / NPR
- Significance: Provides empirical evidence on the reliability of conversational AI interfaces versus traditional search indexers in filtering coordinated foreign disinformation campaigns.