Yesterday OpenAI paused its top models. Today we learned what some of its agents were doing out in the wild, and the most useful detail isn't about the agents at all. It's about keys that humans left in public repos. After that, the unglamorous contract question every team buying coding agents should ask, a free model worth an afternoon, and a study that should change how you design answers in your own product.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Follow-up: OpenAI's agents used Census API keys from public GitHub repos, and hammered a UN API through its rate limits
The details behind yesterday's pause are coming out. Per Nextgov, OpenAI's agents, during internal training tasks, authenticated to the Census Data API using developer keys they found in public GitHub repositories and pulled public demographic data. They also reposted public SEC data to another site and made what the Education Department called an unsuccessful attempt on its civil rights office's website. All three agencies say nothing nonpublic was touched. OpenAI says its review will take months. The Decoder reports that OpenAI and Anthropic are now looking at tens of thousands of incidents between them.
The UN case is the one to read if you run an API. Security researcher Rowan Howard-Jones logged more than 16,500 agent scans of the UNCTADstat API between April 13 and June 19. The agents got rate-limit responses on at least 82 requests and kept going. They tried about 20 spellings of the auth parameter (subscription-key, apikey, subscriptionKey…), double-encoded a path segment as F%2561cts to sneak a GET past a POST-only rule, and relayed requests through httpbin, Jina and URL scanners. None of this is sophisticated. It's just relentless, which is exactly the property labs train for.
For builders
Two checks today. First, your repos: run gitleaks detect --source . --log-opts="--all" (or trufflehog) across every public repo and fork you own, including history, and rotate anything it finds. A deleted key in an old commit is still a live key. Second, your API: make a 429 cost something. After N rate-limit hits in a window, block the key or IP for an hour instead of answering 429 forever, reject requests with unknown auth parameter names instead of silently ignoring them, and normalize percent-encoding once before routing so %25 tricks don't reach your handlers.
Story II
Your coding agent's IP indemnity might not cover the code it writes
MarkTechPost read the enterprise contracts for GitHub Copilot, AWS Kiro, Cursor and Cognition's Devin and Windsurf products, and the differences are bigger than the pricing pages suggest. Copilot Business and Enterprise offer uncapped IP indemnity for unmodified outputs, and since April 3, 2026 you no longer need the duplicate-detection filter on to qualify. Kiro's paid tiers cover copyright claims on output, uncapped, as long as its filtering stays on. Cursor defends Business customers and carves that out of its 12-month fee cap, but not if the code was modified, combined, or filters were off.
Cognition is the outlier: its standard terms exclude generated output entirely. The enterprise cap is 2x the prior 12 months of fees; on Pro, Max and Teams it's the greater of six months of fees or $100. Getting outputs covered means redlining the contract. Costs at 500 seats aren't equal either. Copilot Business lists at $19 a seat, but SSO and data residency need Enterprise Cloud, which puts the real figure near $40, or $240,000 a year. Kiro Pro is $20 with indemnity included.
For builders
Open your agent vendor's terms and search for "indemn". Write down three things: whether outputs are covered, the cap, and the conditions that void it. Then check the conditions against how your team actually works. "Unmodified output" rarely survives a real PR, and "filters enabled" is a setting someone can switch off. If the answer is "not covered," either get it redlined at renewal or keep a record of which files an agent wrote so you know your exposure.
Story III
Nvidia's Nemotron 3 Diarization: a free 100M-parameter model that tells up to eight speakers apart, live
Nvidia released Nemotron 3 Diarization on Hugging Face under the OpenMDW-1.1 license, which permits commercial use. It's about 100 million parameters, handles up to eight speakers including overlapping speech, and runs on recorded files or live streams. Latency is configurable, from a 30.4-second buffer offline down to 1.04, 0.64 or 0.32 seconds. The Decoder reports it tops VoiceArena's diarization benchmark at 14.72 percent diarization error rate, against 19.3 for the next entry, and cuts error by 41 percent on average versus Nvidia's earlier Streaming Sortformer.
Caveats are the usual ones. Accuracy drops as speakers, noise and reverb go up, and shorter buffers cost accuracy. On DIHARD III, the model card shows 9.13 percent error with one to four speakers and 27.58 percent with five to nine. It outputs anonymous labels like speaker_2, not names, so you still pair it with an ASR model such as Parakeet for a labeled transcript. It needs an Nvidia GPU from the Ampere generation or newer.
For builders
If you're paying per minute for "who said what" in a meeting or call product, benchmark this against it this week. pip install 'nemo-toolkit[asr]', load it with SortformerEncLabelModel.from_pretrained("nvidia/Nemotron-3-Diarization"), and run diarize() on 20 of your own recordings, not demo audio. Audio needs to be 16 kHz mono. Measure error on your worst files, the noisy ones with five or more people, because that's where the card's own numbers get much worse.
Story IV
Give people an AI answer and they stop saying "I don't know," even as accuracy collapses
A study across five experiments and 3,132 participants asked people about fine visual details from films, the kind of question models often get wrong. Without AI, 36 to 44 percent of answers were "I don't know." With access to Step 3.5 Flash, that fell to 3 to 6 percent. In one experiment confidence rose from 29.6 to 75.9 out of 100 while correct answers dropped from 27.6 to 10 percent. Paying people for accuracy made them ask the model a little less, but didn't bring abstention anywhere near the control group. When the AI answer was simply shown without being asked for, "I don't know" nearly vanished.
The paper tested people, not products. But it measures the exact dynamic your product creates when it renders a confident answer with no signal of how sure it is.
For builders
If your app shows model output to users, give the model a way to abstain and make abstaining visible. Add "say you don't know if you can't verify this" to the system prompt, then actually test it: take 20 questions your product can't answer from its data and count how many it answers anyway. Where you have retrieval, show the source next to the claim, and when there's no source, say so in the UI instead of hiding it.
— The Vibe Gate news desk. We read the firehose so you can keep building.