$ today_in_10_seconds
OpenAI paused reinforcement learning on its deployment-bound models for two weeks, still has its largest frontier RL run on hold, and for the first time put a number on what supervision costs. Separately, Microsoft patched a Copilot flaw that researchers found by simply asking the assistant why the attack would not work until it told them.
$ lead_story
OpenAI slowed itself down and published the bill
OpenAI's Tuesday post, Pacing model development in an era of cyber-critical capabilities, is the first time the company has said it deliberately slowed scaling on safety grounds. RL training on its latest deployment-bound models was paused for two weeks while it hardened and red-teamed research environments and widened monitoring coverage, and the largest planned frontier RL run remains on hold pending smaller-scale evidence of alignment. The trigger: on August 7 the company concluded that Astra, an unreleased model, may cross the Critical cybersecurity threshold in its Preparedness Framework, the level its own policy says requires stopping. The number to copy into your own planning is this: monitoring overhead runs at roughly 20% of the inference compute being monitored, mandatory across all tool-using RL training and evaluations at Sol capability or above, and across all tool-using inference on Astra.
// why_it_matters: Watching an agent now costs a fifth of running it, a line item nobody's unit economics currently include.
$ also_worth_attention
Copilot talked researchers into its own vulnerability
Varonis disclosed CoSnitch (CVE-2026-24301), a three-flaw chain in Copilot Personal letting one click on a crafted link execute an attacker's prompt and pull data from the victim's connected accounts. Pre-patch, an undocumented autorun parameter paired with the standard query parameter fired the embedded prompt automatically, reaching Gmail, Google Drive, Calendar, Copilot memory and chat history inside the victim's authenticated session. The method is the story: researchers repeatedly asked Copilot to explain why auto-execution was impossible, each refusal mapped more of the architecture, and the assistant eventually volunteered the undocumented parameter unprompted, mid-refusal. Reported December 2025, patched August 18. Treat your model's refusal explanations as attack surface.
link →DeepSeek Harness, one week on
Harness shipped August 13 under MIT alongside V4-Pro-0813, built so that nearly every part of the agent runtime is a swappable plugin, the model included, and drew roughly 27,500 GitHub stars on launch day. Secondary trackers now put it well above 100,000, which measures attention rather than adoption; treat the star count as a snapshot. The consequential change landed August 16: DeepSeek's API price increase took effect, shifting to time-based rates that make off-peak usage cheaper and repeated data retrieval pricier. Free loop, metered model. Schedule your batch agent runs accordingly.
link →River AI raises $1.1B on own the weights
Igor Babuschkin's River AI closed $1.1B led by General Catalyst and AMP PBC, with strategic money from Nvidia and AMD Ventures, roughly four months after incorporating in Nevada. The product is an API for LoRA fine-tuning and reinforcement learning on open-weight models; River claims complex RL runs finish in 15 to 20 minutes with no dedicated infrastructure team, at two to four times the cost savings of closed alternatives, a vendor claim nobody has independently verified. The thesis is that enterprises migrate off off-the-shelf lab models toward privately owned, tuned open weights. The unanswered question is regulatory, not technical: fine-tune the weights and you are unambiguously the deployer of record.
link →
$ bigger_picture
The trend behind the headlines
Three of today's four stories are about the same invisible cost. OpenAI put a percentage on supervision and admitted the biggest run is not running. Varonis showed that the assistant sitting on your connectors will narrate its own architecture to anyone patient enough to keep asking. DeepSeek gives away the harness and meters the model. River is betting a billion dollars that you would rather own the weights than rent them. None of these are capability stories, because capability converged months ago. They are all pricing stories about the layer nobody quotes you: containment, monitoring, and the audit trail. If your agent business case does not have a line for the 20%, it is not a business case yet.
Langfuse
Open-source LLM and agent observability: traces every tool call in a run, attributes token cost per trace and per user, and self-hosts if your data cannot leave your VPC. It will not reduce your monitoring overhead, but you cannot manage a 20% tax you have never measured, and most teams currently cannot tell you what a single completed agent task costs them.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime