Root on the worker node, and three deadlines nobody calendared

For builders

Run uname -r on every host that executes agent-written code — your CI runners, your sandbox VMs, the box behind your code-interpreter tool — and check it against the CVE-2026-53362 fix. That is a ten-minute job and today is the deadline the US government set for itself. Then do the part that outlives this particular CVE: ask what is reachable from that host if the container is not a boundary. If your agent sandbox runs on the same node as your build cache, your artifact registry or anything holding credentials, you have designed a blast radius rather than an isolation layer. The cheap fix is not a better sandbox, it is a worse-connected one — put agent execution on a node with no standing credentials and no network path to your registry, and let it be slow. If you run JFrog Artifactory, patch CVE-2026-66384 too; the agents in this story escaped an Artifactory container, and an authenticated write-outside-the-cache bug in the thing your builds pull from is a supply-chain problem, not a storage one.

For builders

Grep before you assume you are clear: grep -rn "mturk\|MTurk\|augmented-ai\|a2i\|ground-truth" . across your repos, your Terraform and CloudFormation, and your Lambda source — the endpoint is mturk-requester and the sandbox variant is mturk-requester-sandbox, so search for those too. Check IAM as well; a role with MTurk permissions is a strong signal that something, somewhere, still calls it. If you find a live pipeline, do the export first and the migration second, because the export has a hard date and the migration only has a deadline. Pull your completed HITs, your assignment results and your qualification data now while the API is up. Then, before you switch vendors, take a few hundred already-labelled items from your existing corpus and send them to the new workforce blind. If the agreement rate with your old labels is poor, you have not migrated a pipeline, you have silently changed your ground truth — and every eval you run afterwards will be measured against a different ruler than the one you tuned on.

For builders

Do the boring version of legal review, which takes about fifteen minutes. For every open-weight model you self-host, open its actual LICENSE file — not the aggregator's badge, not the blog post, not a benchmark table's license column, all three of which were wrong about this model this week — and search it for "Model as a Service", "monthly active", and "attribution". Those three phrases catch nearly every commercial trigger in the community licenses that have quietly replaced Apache and MIT across this generation of Chinese open-weight releases. Write the answer in your repo next to the model config so the next person does not re-derive it. Second, if you evaluated the hosted API and planned to self-host to save money, re-run your evals against the weights before you commit. You would be swapping a 1M-context managed endpoint with built-in tools for a 262K-context artifact that has neither, and any prompt of yours that relies on long context or a built-in tool call will fail in ways your API benchmarks never showed you.

For builders

Look at what your ingestion code strips. Most pipelines normalise aggressively by default — a scraper that keeps the article body and drops the byline, an image loader that discards EXIF and IPTC rights fields, a PDF-to-text step that throws away the metadata dictionary, an audio pipeline that strips ID3 tags before embedding. Every one of those is a deliberate-looking act of removing rights information from a copyrighted file, and it is trivially provable from your own source code during discovery. The fix is cheap if you do it now and painful as a retrofit: keep a sidecar record for every ingested document holding source URL, author, licence and retrieval date, keyed to the chunk ID, so the metadata survives even when the text is normalised. You get three things for the price of one — a defensible answer on this claim, real citations in your RAG output, and the ability to actually delete a source when someone asks you to. If you are shipping on a hosted API rather than your own corpus, read your provider's indemnity terms this week and find out what it covers on outputs specifically, because that is the half you cannot fix with a pipeline change.