-
Research · WIRED · 14 Oct
WIRED reports Tech Against Terrorism tested 134 AI models with more than 200 prompts tied to terrorist and extremist use, and only two passed completely. The report found abliterated open-weight models now appear online in less than three days on average, down from more than 24 days in 2024.
Why it matters. Open-weight models can lose guardrails within days, raising misuse risk for agent builders and operators that rely on model safety controls.
SafetyModelsPermalinkRead at wired.com
-
Launch · SiliconANGLE · 14 Oct
SiliconANGLE reports SpaceXSI released Grok Voice 3.0 for its Voice Agent API. Artificial Analysis ranked it first on its Speech-to-Speech Index at 86.9, ahead of OpenAI Group PBC’s GPT-Live-1 and Google LLC’s Gemini 3.8 Live.
Why it matters. For teams running voice agents, this adds a same-price model in the Voice Agent API that SiliconANGLE says posted higher benchmark scores than OpenAI and Google.
Agents at workVoiceCompaniesPermalinkRead at siliconangle.com
-
Policy · The Verge · 14 Oct
The Verge reports that eight US states sued OpenAI, alleging it misled customers about ChatGPT's safety and harmed children. OpenAI spokesperson Drew Pusateri said ChatGPT includes protections for different ages and that the company will keep adding safeguards.
Why it matters. It raises pressure to tighten safety claims, underage user protections, and data handling for minors.
SafetyAssistantsCompaniesPermalinkRead at theverge.com
-
Launch · LangChain · 9 Apr
LangChain reports that Deep Agents deploy is in beta as one command for production deployment of an open, model agnostic agent harness with memory, sandboxes, and 30+ endpoints. The post says deployed agents support MCP, A2A, and Agent Protocol, and LangSmith Deployments can be self hosted.
Why it matters. It gives agent teams a beta way to deploy open harnesses with MCP, A2A, and self hosting for memory control.
Agents at workInfrastructurePermalinkRead at langchain.com
-
Research · Huggingface · 14 Oct
Huggingface reports that EviScene is a multimodal ReAct agent for image factuality assessment that investigates images without a supplied textual claim. The paper says EviLens includes 10,174 supported and refuted images, and a trained Qwen3.5-9B agent reached 79.05% balanced accuracy and 52.26% SESR on a 1,527-image test set.
Why it matters. This offers a benchmark and agent setup for checking whether image-based outputs are supported by evidence, not just whether an image is AI-generated.
SafetyModelsPermalinkRead at huggingface.co
-
Research · Hugging Face · 14 Oct
Hugging Face reports a paper that proposes Agent-Reflex, a three-stage framework to train language agents for proactive reflection under noisy tool feedback and user corrections. The paper says its Qwen3-8B version scored 58.6 overall across five long-horizon benchmarks for tool use, plus 29.1 on AgentNoiseBench.
Why it matters. It offers a training approach for agents that must recover from noisy MCP tool feedback and user corrections in real deployments.
ModelsAgents at workPermalinkRead at huggingface.co
-
Analysis · SiliconANGLE · 13 Oct
SiliconANGLE reports that Spranic, based in Munich, is creating a spatial memory layer to record physical spaces and their changes. Founder and chief executive officer Jordi Spranger said linking that data with knowledge graphs could help robots interpret environments across tasks and make decisions auditable.
Why it matters. Reusable spatial context and knowledge graphs could reduce task-specific planning and give robots auditable information sources across different tasks.
SafetyInfrastructureCompaniesPermalinkRead at siliconangle.com
-
Money · SiliconANGLE · 14 Oct
SiliconANGLE reports Evos raised $9 million in pre-seed funding for Workhorse, an agentic operations platform for smaller industrial businesses. Founder Urav Shah said its agents learn tasks behind an approval gate and work autonomously only after proving themselves with human oversight.
Why it matters. For builders, it highlights agents that learn behind an approval gate and move to autonomous work only after proving themselves with human oversight.
Agents at workCompaniesPermalinkRead at siliconangle.com
-
Money · SiliconANGLE · 14 Oct
SiliconANGLE reports that Uvi raised $7.5 million in seed funding led by CT Innovations, with Cardumen, Mana VC and 91 Ventures. Uvi says its benchmarks show internally built AI agents cost about eight times more than they should without its context reasoning layer.
Why it matters. For teams running AI agents, the pitch is lower per-response costs and less repeated relearning of company context across tools and data.
Agents at workInfrastructureCompaniesPermalinkRead at siliconangle.com
-
Analysis · TechCrunch · 13 Oct
TechCrunch reports that Laplace is building automated red-teaming for robotics and physical AI using a high-fidelity simulation and AI agents to probe failures before deployment. CEO Howie Sun said the startup has revenue and customers including Nvidia.
Why it matters. It points to automated red-teaming for physical AI, which could help teams find edge cases in robots and autonomous vehicles before deployment.
SafetyInfrastructureCompaniesPermalinkRead at techcrunch.com
-
Launch · TechCrunch · 14 Oct
TechCrunch reports that Apate.ai deploys thousands of AI agents that pose as victims across calls, messages, email, and the web to collect fraud intelligence. The startup says customers including Commonwealth Bank of Australia and TPG Telecom use that data to inform anti-fraud measures.
Why it matters. It shows agent fleets can be used not just for workflows, but to engage scammers and turn those interactions into anti-fraud intelligence for banks, telecom companies, and government agencies.
SafetyAgents at workCompaniesPermalinkRead at techcrunch.com
-
Launch · The Decoder · 14 Oct
THE DECODER reports that Artemis Security launched Orion-1, a model for investigating and responding to attacks on corporate networks. Artemis says it links logs across systems, scored 80.8 in attack reconstruction versus 58.3 for the next-best model, and is in private preview.
Why it matters. A purpose-built security model may avoid guardrails that blocked defenders and let teams automate some network attack analysis and response with human approval for high-impact actions.
SafetyInfrastructureCompaniesPermalinkRead at the-decoder.com