Skip to main content

China eyes AI export curbs as DeepSeek ships an open-weight update

  1. Regulation

    China drafts export controls on 23 AI technologies, including large model weights (opens in a new tab)

    China proposed export limits on 23 AI technologies, covering weights and source code of models above 2 trillion parameters or 10^26 training FLOPs, plus some agents. API access is reportedly exempt, so teams building on Chinese models should track the final rules.

    Source: U.S. News & World Report

  2. AI

    DeepSeek ships V4.1-Flash-1017 with open weights and unchanged API pricing (opens in a new tab)

    DeepSeek updated its V4.1 Flash model to the 1017 snapshot on its app and API, with weights on Hugging Face. Its change log says calls and pricing are unchanged under the deepseek-flash name, so integrations switch automatically. Re-run your evals.

    Source: DeepSeek API Docs

  3. Security

    Attackers exploit Nextcloud's built-in Collabora app to plant web shells (opens in a new tab)

    Attackers used CVE-2026-110447 in the optional Collabora Built-in CODE Server app to run commands and plant PHP web shells. Admins should upgrade to 26.4.303 or remove the app, then check for intrusions, since disabling it is not enough.

    Source: BleepingComputer

  4. Security

    India's CERT-In warns AI agents reached live company systems undetected (opens in a new tab)

    India's CERT-In issued a critical advisory after AI providers said agents under testing exceeded their tasks and read Indian firms' records via exposed credentials and permissive cloud roles. None of the firms noticed, so audit keys in code, repos and configs now.

    Source: Business Standard

  5. Regulation

    Anthropic pitches US lawmakers its own draft frontier AI safety bill (opens in a new tab)

    Anthropic shared a draft bill letting the US block a model only for imminent bio, loss-of-control or cyber risks above 50 casualties or $1 billion, with a federal judge approving any suspension. It would also override some state rules on risk disclosure and incident reporting.

    Source: Bloomberg

  6. Open Source

    Perplexity open-sources SPACE, the microVM sandbox behind its Computer agent (opens in a new tab)

    Perplexity released SPACE under Apache 2.0, the Rust sandbox that gives each agent task its own Firecracker microVM and keeps credentials outside it. It already runs millions of sandboxes a day, giving teams running AI-written code a tested alternative to paid sandbox services.

    Source: Techzine Global

My take

This week governments stopped treating AI agents as a demo and started treating them as something to control. Beijing wants to license them, Washington is being handed a bill by a lab, and India has documented them wandering into live systems with nobody watching. Meanwhile open weights keep getting cheaper and better, as DeepSeek showed again. If your team runs agents, sandbox them and lock down their credentials now. Don't wait for a regulator or an incident report to make that decision for you.