JadidJust-in AI Developments, Incidents & Deadlines
October 2026

AI news, Saturday, October 3, 2026

9 stories

TLDR

  • Critical GitLab AI Gateway flaw (CVE-2026-90970, CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 9.9) patched — self-hosted Duo Agent Platform sandbox escapeBreaking out of a sandbox, the restricted environment that limits what untrusted code can reach.NIST ↗ (Oct 2) Security
  • "PixelLeak": coding agents exposed 13,000+ internal screenshots from 300+ orgs via public repos (Sep 29) Security Agents
  • Salesforce signs ~$2B deal for Listen Labs (AI customer-research agents); Nebius buys Inferize for GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ cold-start tech (Oct 1) Customer AI Infra
  • Meta publishes six math papers co-developed with Muse Spark; five answer open questions (Oct 2) Research
  • Trump reportedly set to name DNI Jay Clayton as AI czar (Oct 2) — expected, not yet formally announced Policy

Action Items

Things your engineering team should look at or do this week.

  • Patch self-hosted GitLab AI Gateway to 19.2.4 / 19.3.2 / 19.4.1 if you run Duo Agent Platform (vulnerable: 18.1.6–19.4.0). GitLab.com and GitLab-hosted gateways are not affected.
  • Add guardrails against agents creating public repos — block new public repositories, personal-account pushes and visibility changes from agent credentials; remove blanket auto-approval in agent configs.
  • Scan org members' personal GitHub accounts (93% of PixelLeak exposures were in employee personal repos) for stray agent-created screenshot repos.

Models & Tools

  • DeepSeek Harness v0.2 public preview (Oct 2 coverage; announced Sep 29) Agents Open Source — Desktop installers for macOS (Apple silicon) and Windows; plugin manager, experimental plugin-creation-by-chat mode, scheduled automation tasks, six official plugins (Agent Teams, Shell, Subagent, Web Search, etc.). Update to the Harness items covered Sep 26. (Report: MPost)
  • Cactus Whistle (Oct 2) Open Source — 16.9 MB on-device speech-to-text model; beats Whisper base (145 MB) on LibriSpeech, SPGISpeech and Earnings-22 but trails on TED-LIUM, AMI and MLS. Weights on Hugging Face, pip install cactus-needle; runs CPU-only across 17 targets. (Cactus blog)
  • Nebius acquires Inferize (Oct 1) Infra — Israeli 10-month-old startup (ex-Granulate founders, 17 staff) that cuts GPU idle time from model-weight cold starts; reported $100–150M, terms undisclosed. (Calcalist)

Security

  • GitLab AI Gateway CVE-2026-90970 (Oct 2) Agents — CVSS 9.9; authenticated Duo Agent Platform users can escape the prompt-template sandbox via crafted flow config and run arbitrary commands on self-hosted gateways. No known exploitation or public PoCProof of conceptPublic exploit code that shows a vulnerability can actually be triggered.FIRST ↗. (Security Affairs)
  • PixelLeak (Sep 29) Agents — Agents unable to attach images to PRs from the CLI created public repos to host screenshots, exposing 13,000+ internal images across 900+ repos at 300+ organizations (incl. Fortune 500s and AI labs); open-source gitshot tool involved in ~1/3 of cases. (Glow)

Industry & Business

  • Salesforce to acquire Listen Labs (Oct 1) Customer AI — ~$2B (not officially disclosed); AI agents that recruit, interview and simulate customers; to be folded into Marketing Cloud and Service Cloud; expected close Q4 FY2027 pending regulatory approval. (SiliconANGLE)
  • AMD to acquire World Labs for $8.2B (Sep 28) Infra — Fei-Fei Li's world-model startup joins AMD; Li becomes chief scientist. Announced, not yet closed. (CNBC)

Policy & Safety

  • Jay Clayton expected to be named White House AI czar (Oct 2) Policy — Reported by CNN, Bloomberg, NBC and others: the DNI would take the role while keeping his intelligence post, amid rising AI safety concerns. Not yet officially announced. (CNN)

Paper Spotlight

  • Meta + mathematicians: six papers with Muse Spark (Oct 2) — Using Muse Spark 1.1/1.2 in Thinking Mode through the regular meta.ai chat (no custom scaffold), collaborators resolved five open questions, including a disproved 2024 group-theory conjecture and a sharp threshold for fitting random Gaussian points to ellipsoids. (Meta Research)

Type to search · ↑↓ to move · enter to open