Monday, October 5, 2026

TLDR

Q3 2026 global venture funding hit $159B, with AI at 64% and a record 27 billion-dollar rounds (Oct 5) Trending
Collibra acquires Munich-based trail ML to add runtime enforcement to AI governance (Oct 5) Agents
Mistral Vibe CLI permission-bypass CVEsCommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ (CVE-2026-87987/87984) enable code execution via malicious repos; fixed in 2.25.8+ (Sep 23–24) Security Agents
OpenCode localhost API flaw lets a webpage trigger RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ via npm install; fixed in 1.18.22 (Sep 24) Security

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

Audit coding-agent CLI versions across the team: Mistral Vibe ≥ 2.25.8, OpenCode ≥ 1.18.22, Codex CLI ≥ 0.149.0, Codex Desktop ≥ build 26.818.21641.
Treat local agent APIs as an attack surface: confirm any agent tool you run exposes no unauthenticated localhost endpoints (OpenCode's 127.0.0.1:4096 case), and avoid opening untrusted repos in agents with auto-approve.
If you evaluate AI governance tooling, note the runtime-enforcement trend (Collibra + trail ML) and ask vendors whether policy checks block actions pre-execution.

Mistral Vibe permission bypass: CVE-2026-87987, CVE-2026-87984 (plus -87985/-87986/-87988) (Sep 23–24) Agents — The Bash tool's Tree-sitter permission check saw a simplified command while the shell executed the full string (env assignments, redirects), so repo-controlled content could run arbitrary code and exfiltrate API tokens. Patched starting 2.25.4; update to 2.25.8+. (SecMate)
OpenCode RCE via upgrade endpoint, GHSA-632h-h47v-g4x4 (Sep 24) — A malicious webpage could POST an HTML form to the local API (127.0.0.1:4096) and make /global/upgrade install an attacker npm package with a preinstall script. Affects 1.14.30–1.18.21 (npm/pnpm/Bun installs); fixed in 1.18.22. Datadog notes ~647k downloads of vulnerable versions in mid-September. (Datadog Security Labs)
OpenAI Codex sandbox escapesBreaking out of a sandbox, the restricted environment that limits what untrusted code can reach.NIST ↗ "Overpatch" and "Heapjack" (Sep 15) — apply_patch could widen write access to the whole disk without prompts, and Codex Desktop leaked a trust tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ readable from heap snapshots. Both reported Aug 12 and patched within eight days; update to CLI 0.149.0+ / Desktop 26.818.21641+. (Accomplish)

Global venture funding Q3 2026: $159B (Oct 5) Trending — Down 25% from Q2 but up 53% YoY. AI startups took $102B (64%); a record 27 rounds topped $1B, led by Databricks and Safe Superintelligence at $5B each. (Crunchbase News)
Collibra acquires trail ML (Oct 5) Agents — Adds agent-driven policy automation, continuous control assessment and runtime blocking of policy-violating actions, mapped to the EU AI Act, ISO 42001 and NIST AI RMF. Terms not disclosed in the coverage. (Unite.AI)

Trump order renames "AI" to "super intelligence" in federal usage and launches America.gov (Sep 29) Policy — Executive order creating an AI-powered federal services portal. Mostly terminology and portal plumbing; watch for downstream procurement language. (White House)