Tuesday, June 23, 2026

The big new signals: OpenAI shipped its defensive cyber model and an open-source patching program, SpaceX landed another multi-billion compute customer, the Five Eyes issued a rare joint AI-cyber warning, and Claude took a second broad outage in under 14 hours. One fresh, actionable dev-tool CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ (Cursor) to patch now.

TLDR

OpenAI fully releases GPT-5.5-Cyber and launches "Patch the Planet" (Jun 22) — a defensive cyber model (85.6% on CyberGym) plus a Trail of Bits/HackerOne program to find-and-fix bugs in 30+ critical OSS projects (cURL, Go, Python, aiohttp) Security
SpaceX signs up to $6.3B compute deal with open-source lab Reflection AI (Jun 22) — $150M/month for GB300 capacity at Colossus 2, joining Anthropic, Google, and Cursor as Colossus customers Infra Trending
Five Eyes agencies issue rare joint warning: AI-supercharged cyberattacks are "months, not years" away (Jun 22) — CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗/NSA + UK/CA/AU/NZ urge whole-of-org defensive prep Policy Security
Claude hit by a SECOND broad outage in under 14 hours (Jun 23) — all models/platforms except Claude for Government; resolved by ~16:44 UTC Production
Cursor RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ via allowlist bypass (CVE-2026-22708) — shell built-ins escape Auto-Run allowlist; patch to v2.3+ Security

Action Items

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

Update Cursor to v2.3 or later now. CVE-2026-22708 lets indirect prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ bypass the Auto-Run allowlist via shell built-ins (export, env-var poisoning of PATH/LD_PRELOAD), turning trusted commands like git/npm into RCE. Audit any agent running in allowlist/auto-run mode. (SC Media)
Re-check Claude failover after two outages in 14 hours. Both the Jun 22 and Jun 23 incidents hit every model at once across API/Code/Cowork — single-model fallback isn't enough. Confirm cross-provider escape hatches and backoff for customer-facing paths. (TechRadar)
Security leads: read the Five Eyes advisory and evaluate GPT-5.5-Cyber / Patch the Planet for defensive use. If you maintain or depend on critical OSS, consider the patching program; prioritize foundational controls per the advisory. (Cyber.gov.au | OpenAI)

Models & Tools

OpenAI ships GPT-5.5-Cyber and "Patch the Planet" (Jun 22) Security Open Source — OpenAI expanded its Daybreak program with the full release of GPT-5.5-Cyber, its most capable defensive security model (a new 85.6% on the CyberGym benchmark), plus an updated Codex Security plugin and a partner program. The headline initiative, Patch the Planet (built with Trail of Bits and HackerOne), pairs AI-assisted vulnerability research with mandatory human expert review; an initial five-day sprint surfaced hundreds of issues and merged dozens of patches across 30+ projects that have committed to participate, including cURL, Go, Python, Sigstore, pyca/cryptography, aiohttp, and freenginx. (OpenAI | SiliconANGLE | The Hacker News)

APIs & Pricing

DeepSeek retires legacy API aliases after Jul 24 (deadline) Pricing — The deepseek-chat and deepseek-reasoner aliases are fully retired after Jul 24, 2026 15:59 UTC; migrate to the V4 model IDs (deepseek-v4-pro, deepseek-v4-flash, 1M context, MIT-licensed, shipped Apr 24). Anyone pinned to the old aliases should update calls before the cutover. (DeepSeek API Docs)

Security

Cursor allowlist bypass enables stealthy RCE (CVE-2026-22708) (Jun) Security — Before v2.3, with Auto-Run + Allowlist enabled, shell built-ins (export, unset, set) execute without approval and can poison environment variables (PATH, LD_PRELOAD) so a trusted allowlisted command resolves to an attacker binary. Triggerable via indirect prompt injection from poisoned files, repos, or web content. Fix: update to v2.3+, which now requires explicit approval for commands the parser can't classify. (SentinelOne | Pillar Security)

Industry & Business

SpaceX signs up to $6.3B compute deal with Reflection AI (Jun 22) Infra Trending — The open-source lab (last valued at ~$25B) will pay ~$150M/month from Jul 1 for Nvidia GB300 capacity at Musk's Colossus 2 data center, totaling ~$6.3B if it runs through 2029 (either side can exit on 90 days' notice after month three). It cements Colossus as a commercial compute platform alongside Anthropic, Google, and Cursor — and adds a strategically open-source customer. (CNBC | TechCrunch)
Samsung rolls out ChatGPT and Codex to employees (Jun 22) Customer AI — Samsung Electronics deployed OpenAI's ChatGPT and Codex internally, an enterprise-adoption signal landing the same day as OpenAI's security push. (Build Fast with AI)

AI in Production

Claude suffers a second broad outage in under 14 hours (Jun 23) Production Trending — Update to the Jun 22 incident. Anthropic began investigating elevated error rates at 14:19 UTC, identified the issue by 14:25, and marked it resolved by ~16:44 UTC. Downdetector peaked at ~7,100 US reports. It hit all models and all platforms — Claude.ai, Console, API, Code, and Cowork — except Claude for Government, making it broader than the prior night's outage. (Some outlets noted unverified third-party claims of a data leak that Anthropic said it was reviewing — treat as unconfirmed.) The takeaway for production traffic: two all-model outages inside one window argue for cross-provider redundancy, not just model fallback. (TechRadar | Status)

Policy & Safety

Five Eyes issue rare joint warning on AI-driven cyber threats (Jun 22) Policy Security — Cyber chiefs from the US (CISA, NSA), UK (NCSC), Canada, Australia, and New Zealand released a joint advisory warning that next-generation frontier models will supercharge offensive hacking "on a timeline measured in months, not years." The guidance urges leaders to assess risk and accountability, prioritize foundational controls, empower cyber leaders, and stay engaged as the threat evolves — directly echoing the rationale behind recent US export controls on Anthropic's Fable 5 / Mythos 5. (Cyber.gov.au | The Hill)

Busier cycle than the weekend.