LastWeekIn.Tech

Today's AI news centers on security incidents involving AI agents and API exposure, a major geopolitical move on AI export controls from China, and notable model/architecture research from Mistral and the open research community.

1. 1

   Number 1. Policy & Regulation aiimpacthub.com

   China proposes export controls covering 23 AI technology categories

   China's commerce and science ministries opened a public consultation on October 16 proposing to add 40 technologies to its export-control catalogue, including 23 AI-related categories. Comments on the proposal close November 15, while a separate interim registration requirement for covered exports is already in effect. Technology already made public through open-source releases or academic papers is excluded from the proposed controls.

   Why it matters This signals Beijing tightening its grip on AI technology flows even as it pushes open-source releases domestically, mirroring the export-control tit-for-tat already underway with the US on chips. The carve-out for open-source and academic work is notable, potentially protecting China's open-weight model ecosystem (e.g., DeepSeek) while restricting proprietary tech transfer. Companies with China-linked AI supply chains or research partnerships now face a narrowing compliance window before the rules solidify.

   What to watch Watch for the final rule text after the November 15 comment period closes, and for how multinational AI firms adjust technology-sharing agreements with Chinese partners in response.

2. 2

   Number 2. Security & Safety The Verge

   A hacker broke into John Deere and SharkNinja with an army of AI agents

   A hacker reportedly breached more than 100 companies' websites, including John Deere, SharkNinja, and Cummins, using a largely autonomous army of AI agents, according to findings from security firm Norforge reported by Reuters. The AI agents carried out most of the attack sequence with minimal human involvement from the hacker.

   Why it matters This is one of the clearest documented cases of AI agents being used at scale to automate real-world intrusion campaigns against major industrial and consumer companies, rather than isolated proof-of-concept demos. It validates growing concern among security researchers that offensive AI agents can dramatically lower the skill and effort bar for large-scale breaches, compressing attacker timelines from weeks to hours and complicating detection since human oversight is minimal.

   What to watch Expect incident response disclosures from the named victims and scrutiny of which AI platforms or frameworks were exploited, plus possible regulatory or industry calls for agent-usage monitoring standards.

3. 3

   Number 3. Security & Safety The Hacker News

   OpenCode Console Breach May Have Exposed OpenAI, Anthropic, and Google API Keys

   OpenCode disclosed that an intruder accessed a legacy system behind OpenCode Console on the night of October 16, potentially exposing third-party API keys—including OpenAI, Anthropic, and Google credentials—that users had stored there. The affected group represents 1.4% of Console users who had linked an outside provider key, and OpenCode has emailed them instructing immediate key revocation and replacement.

   Why it matters Beyond the earlier OpenCode credential leak already reported, this incident specifically implicates stored keys for the three largest frontier-model providers, meaning compromised keys could let attackers rack up usage charges or access account-linked data on OpenAI, Anthropic, and Google infrastructure. It underscores a recurring weak point in the AI tooling ecosystem: third-party wrappers and consoles that aggregate provider credentials become high-value targets.

   What to watch Watch whether OpenAI, Anthropic, or Google report anomalous usage tied to the leaked keys, and whether other AI developer-tool platforms tighten how they store and rotate customer provider credentials.

4. 4

   Number 4. Capabilities Hacker News

   Hybrid Reasoning Agents: LLM agents that think in latent space, act through text

   Researchers published a paper on "Hybrid Reasoning Agents" (HybridAgents), a framework for LLM agents that can dynamically switch between reasoning in latent space and interacting with tools and environments through text. The approach aims to combine the efficiency of latent internal reasoning with the transparency and tool-access benefits of text-based reasoning.

   Why it matters Latent-space reasoning has been pitched as a way to make agents faster and more computationally efficient than token-by-token chain-of-thought, but it sacrifices interpretability and the ability to call external tools mid-thought. A hybrid architecture that lets agents move fluidly between the two modes could meaningfully improve agent efficiency without losing the auditability regulators and safety researchers increasingly demand.

   What to watch Watch for benchmark comparisons against standard text-reasoning agents on tool-use and long-horizon tasks, and whether major labs (OpenAI, Anthropic, Google) adopt similar hybrid-reasoning techniques in production agents.

5. 5

   Number 5. Open Source Hacker News

   Timestral: Mistral release time-series foundation models

   Mistral released Timestral, its first family of time-series foundation models, available in four sizes ranging from 5 million to 209 million parameters. The models deliver probabilistic, multivariate forecasts, support past and future covariates, and are designed to run efficiently on CPUs while achieving strong zero-shot forecasting performance against larger competing models.

   Why it matters Time-series forecasting is a massive but underserved application area compared to text and image generation, spanning demand forecasting, finance, and industrial monitoring. By releasing a lightweight, CPU-efficient foundation model family that reportedly beats larger models on key benchmarks, Mistral is pushing into a new vertical and reinforcing its open-model strategy against closed, narrower forecasting tools from legacy vendors.

   What to watch Watch for independent benchmarking of Timestral against established forecasting libraries and models, and whether Mistral expands the family with larger or domain-specific variants.