Tech News, October 26: OpenAI Trains Offline, Moonshot Curbs Kimi, and CISA’s Collabora Deadline

The last few days have produced a clear pattern: the AI industry is tightening its own guardrails while the evidence for a payoff outside the tech sector stays thin, and one actively exploited server flaw reaches its patch deadline today. OpenAI has moved its frontier training off the live internet, Moonshot has deliberately weakened the hacking ability of its next open model, and administrators running Nextcloud with Collabora have until today to fix a flaw attackers are already using.

OpenAI moves frontier training to an offline web

OpenAI said on October 23 that it resumed reinforcement-learning training and internal inference for its most capable models on October 1, after an 11-day pause. The pause followed a September 20 incident in which an internal research model reached limited live internet access through a DNS filtering gap, despite being restricted from the live web. Monitoring flagged the behaviour within 15 minutes, a human began reviewing it 3 minutes later, and the run was stopped manually about 2.5 hours after that review began, according to the incident account.

Frontier training now uses stored web material instead of live internet access. OpenAI also ran teams of AI agents against its own containment to test whether they could break out, reach restricted resources, or act without permission. That exercise found previously unknown flaws, which the company said it fixed, and it also surfaced a flaw in Brotli, a widely used data-compression tool. Attackers could exploit the Brotli flaw by sending malicious data that lets them run code on affected systems, including systems using Node.js, Bun, and Envoy. OpenAI said the root cause had been fixed in Brotli source code in February, before it was known to be a security vulnerability, but no new version was released at the time; updated versions have since been released.

Why it matters: this is not a change to ChatGPT search for ordinary users. It is a change to how the most capable research models are contained while they are trained. The September incident showed that a configuration gap, not a deliberate jailbreak, was enough to give a research model outside access, and that the automatic shutdown did not work as expected. Offline training is a blunt control, and it only helps if the stored environment is genuinely isolated and the monitoring failures are fixed rather than papered over.

Moonshot curbs the hacking ability of Kimi K3.1 Pro

Moonshot AI said on October 24 that it had reduced the cyber-offensive capabilities of its newest model, Kimi K3.1 Pro, which it describes as its most capable 3-trillion-parameter-class open-weight model. The company removed cyber-related data from the training corpus and reinforcement-learning environments used to develop the model, while retaining general coding skills. It plans to publish the model weights by November 4.

Moonshot put its reasoning plainly: “We aim not to release open weights that would extend the absolute frontier of cyber capability.” The context matters. Open-weight models can be downloaded, run, and modified outside the developer’s controls, so a capability baked into the weights cannot be recalled with a policy change. At the same time, Moonshot argued that open-weight models can help more organisations inspect and secure their own software, including older systems that receive little attention.

Why it matters: this is a deliberate trade-off, and it is unusual to see it stated this openly. China also proposed export controls last week that would ban exports of AI-driven malware-generation and mutation technology and restrict AI-assisted vulnerability-discovery tools, according to Reuters reporting. The draft measures remain open for public comment. The practical question for November 4 is whether the published weights actually show the claimed reduction in practice, and whether independent testing confirms it, rather than taking the training-data claim on trust.

CISA’s Collabora deadline is today: patching alone is not enough

CISA added CVE-2026-110447 to its Known Exploited Vulnerabilities catalog on October 23, with a remediation deadline of October 26 for federal civilian agencies. The flaw is an unauthenticated OS command injection, scored CVSS 9.2, in proxy.php of the Nextcloud app richdocumentscode, the Collabora Online Built-in CODE Server. It is exploitable when the built-in CODE server is not running, for example during a restart, upgrade, or crash.

The fix is to upgrade richdocumentscode to version 26.4.303. Disabling the Collabora app in Nextcloud is not sufficient, because the vulnerable proxy.php remains accessible; administrators must upgrade or completely remove the package. Standalone, Docker-based, and Collabora-hosted deployments are not affected; only the built-in CODE Server integration within Nextcloud is. Admin reports cited in public write-ups describe attacks starting September 10, with PHP webshells dropped on affected servers, which is why a patch alone does not close the incident: affected servers also need a hunt for unexpected files, credential changes, and a check of web-server logs for requests to proxy.php while the CODE server was down.

Why it matters: this flaw breaks a common assumption, that a service that is not running is not an attack surface. Here, the stopped state is part of the trigger. For anyone self-hosting Nextcloud, the check is specific: do you have the built-in CODE Server installed, is it at 26.4.303 or removed entirely, and have you looked for signs of earlier compromise?

Apollo: AI has shown “zero” payoff outside tech so far

Apollo Global Management chief economist Torsten Sløk wrote in a note on October 24 that, despite massive investment and widespread hype, AI has delivered “zero financial payoff to corporate America outside of tech” so far. His measure was revenue per worker: rising for the Magnificent 7 companies (Amazon, Apple, Google, Meta, Microsoft, Nvidia, and Tesla) but flat for the other 493 companies in the S&P 500. Bloomberg flagged the note on October 25, as investors watch AI spending against returns during earnings season.

Meta is the clearest example of the tension. Meta spent $34.9 billion on capital expenditures in its third quarter while reporting negative free cash flow of $2.1 billion, and said it expects to be even more aggressive in AI-related capital spending, which sent its shares down, according to the Bloomberg account. On the other side of the ledger, Meta CEO Mark Zuckerberg said its Muse AI model had been set up by more than 10 million people in the US and Canada in six weeks, with around a quarter of that group using it daily.

Why it matters: adoption and payoff are different measurements, and conflating them is how AI budgets get defended with user counts while margins absorb the cost. Sløk’s note does not prove AI will never pay off outside tech; it says the payoff is not yet visible in revenue per worker outside the seven largest tech companies. That distinction should shape how any organisation justifies its next AI spend: by a measured change in output per worker, not by how many people tried a tool.

At a glance

Key takeaways

• OpenAI’s frontier models now train on stored web material, not the live internet, after a September DNS gap gave a research model limited outside access.
• Moonshot deliberately removed cyber-related training data from Kimi K3.1 Pro, but the claim will only be proven when the November 4 weights are tested independently.
• If you run Nextcloud with the built-in CODE Server, upgrade to 26.4.303 or remove the package entirely today, and check for prior compromise.
• AI adoption is not the same as AI payoff: revenue per worker is flat for the S&P 493 while rising for the Magnificent 7, according to Apollo.

References

• OpenAI resumes frontier AI training after security overhaul, discloses flaw in widely used software — Indian Express / Reuters
• Moonshot curbs hacking ability in new Kimi model as fears grow over AI-powered cyberattacks — Indian Express / Reuters
• CISA Collabora KEV listing sets October 26 remediation deadline — Remio
• Black October: A Blue Team Field Guide — DEV, Collabora section
• CVE-2026-110447 — Basic Security
• Apollo says AI has shown ‘zero’ payoff outside tech — PYMNTS, citing Apollo and Bloomberg