-
Incident · OpenAI · 30 Sep
OpenAI says it broke up a campaign aimed at extracting protected model reasoning and is bolstering defenses against adversarial distillation. OpenAI said people linked to Moonshot AI were at the core of the effort, which began in the first week of July.
Why it matters. It points to bolstered defenses against model distillation and extraction of protected reasoning from GPT systems.
SafetyModelsCompaniesPermalinkRead at openai.com
-
Launch · Cloudflare Blog · 30 Sep
Cloudflare Blog reports that more than half of Internet traffic is now non-human, and AI agent requests on Cloudflare's network rose by more than 1,700% over the past year. It says Search, Agent, and Training controls and Pay Per Use are meant to help sites set terms and get paid.
Why it matters. If agent traffic keeps rising, builders may need separate rules for search, agents, and training, and a way to charge when content is used.
Agents at workInfrastructureCompaniesPermalinkRead at blog.cloudflare.com
-
Policy · CNBC · 30 Sep
CNBC reports California enacted SB 947 barring employers from relying solely on AI for firing or discipline and requiring human corroboration when AI is used primarily. CNBC reports workers must get written notice, the data description, and a human contact.
Why it matters. Teams using workplace AI in California now need human review, worker notices, and data descriptions for termination or discipline decisions.
Agents at workCompaniesPermalinkRead at cnbc.com
-
Launch · Amazon Web Services · 29 Sep
AWS reports Amazon Bedrock now offers in-region inference in Seoul for Claude Opus 5 plus Claude Sonnet 5, and in Singapore for Claude Sonnet 5. AWS says inference requests and data stay in the AWS Region called through the bedrock-runtime endpoint.
Why it matters. This helps applications in South Korea or Singapore meet strict data residency requirements by keeping prompts and results in one AWS Region.
ModelsInfrastructurePermalinkRead at aws.amazon.com
-
Launch · GitHub · 12 Jun
GitHub reports Magnitude is an open source engine for agents that tunes kernels on each device before use. The repo says it reaches up to 2x the speed of llama.cpp and works on Apple Silicon, NVIDIA, AMD, or CPU.
Why it matters. This gives agent builders a local engine that connects to Pi, OpenCode, Hermes, Codex, and more while keeping prompts, files, and models on their machine.
ModelsAgents at workInfrastructurePermalinkRead at github.com
-
Research · Huggingface · 30 Sep
Huggingface reports a paper on AsyncLLM, an open-source system for concurrent observing, reasoning, and acting by pretrained LLMs without extra training. The paper says Qwen 3.x worked asynchronously on video, videogames, and monitoring with no task-specific training.
Why it matters. This points to agents that can handle overlapping work instead of only sequential cycles.
ModelsAssistantsInfrastructurePermalinkRead at huggingface.co
-
Policy · SiliconANGLE · 29 Sep
SiliconANGLE reports that Equals Money lets customers’ AI tools use its MCP server for read-only data access and nonsensitive writes, not payment instructions. At Okta’s Oktane event, James Simcox said the company treats agents like staff with identities, auditing and token revocation.
Why it matters. For teams exposing MCP servers to agents, this favors keeping payment actions outside agent access and using identity and token revocation controls.
SafetyInfrastructureCompaniesPermalinkRead at siliconangle.com
-
Money · SiliconANGLE · 30 Sep
SiliconANGLE reports Metaview raised $60 million in a Series C led by Insight Partners. Metaview said fillmore, a recruiting agent in Slack, is generally available, and more than 7,000 companies already use its tools.
Why it matters. It shows a live use case for Slack based agents that automate recruiting work such as outreach and scheduling.
Agents at workCompaniesPermalinkRead at siliconangle.com
-
Money · TechCrunch · 30 Sep
TechCrunch reports that Flow Engineering raised a $50 million Series B at a $750 million valuation. The startup says its AI agents help align CAD drawings with product requirements, simulations, and testing for hardware design.
Why it matters. It signals more backing for AI agents used in hardware design work.
Agents at workCompaniesPermalinkRead at techcrunch.com
-
Launch · Vercel · 30 Sep
Vercel reports that Sandbox now works with Secure Compute, letting sandboxes use a team's dedicated network and static IPs for public internet traffic. It also says sandboxes can reach private AWS VPC resources through VPC peering, and the feature is available to Enterprise teams with Secure Compute.
Why it matters. Teams can run sandboxes on a dedicated network with static IPs and VPC access, which can simplify agent connections to private AWS resources.
InfrastructureCompaniesPermalinkRead at vercel.com
-
Money · TechCrunch · 30 Sep
TechCrunch reports Restate raised a $20 million Series A; Singular led it, and Redpoint Ventures plus Capital One Ventures also invested. TechCrunch says interest in Restate's durable workflow infrastructure has risen as AI agents increase the need for resilient, reproducible processes.
Why it matters. It signals more investment in infrastructure for AI agents that need durable, reproducible workflow execution.
Agents at workInfrastructureCompaniesPermalinkRead at techcrunch.com
-
Research · Hugging Face · 30 Sep
Hugging Face reports a paper on on-policy distillation found an early useful-transfer phase where held-out accuracy rose about linearly with reverse KL from student initialization. The paper says each observed weak-to-strong setup let the student's peak gold score top the teacher's own.
Why it matters. For AI agent builders, the results suggest a smaller RL expert can supervise a larger model, and teacher score alone does not define supervision value.
ModelsPermalinkRead at huggingface.co