5-min read · Curated daily by an AI Systems Architect
Focus: OpenAI Codex Expansion, Agent Benchmarks, AI Supply Chain Security
1. 🔗 OpenAI Acquires Ona to Strengthen Codex Cloud Runtimes
OpenAI announced the acquisition of Ona, a startup that builds secure, pre-configured cloud runtime environments for AI agents. The deal — terms undisclosed — will fold Ona's entire team into the Codex division.
Why it matters: Codex now has 5M+ weekly active users (up from 3M in April), and this acquisition directly addresses the "long-running task" gap. Ona's technology gives AI agents secure access to tools, systems, and context data needed for production-grade autonomous work. This is OpenAI's third coding-tool acquisition in 12 months, following Astral (Python tooling) and Alex Codes (Xcode plugin).
2. 🔗 Agents' Last Exam: Codex + GPT-5.5 Scores 8.6%, Claude Code Gets Zero
UC Berkeley and 250+ industry experts released the Agents' Last Exam (ALE) — a brutal new benchmark of 1,490 real-world professional tasks spanning manufacturing, law, medicine, and visual media.
The results are sobering:
- Codex + GPT-5.5: 8.6% pass rate (best in class)
- Claude Code + Opus 4.7: 0% on the hardest tier
- Average mainstream system: 2.6%
- 78% of failures stem from understanding/method gaps, not execution
The benchmark auto-grades 93.2% of tasks, making it scalable. The research team's key finding: model choice has ~3x the impact of agent framework choice on real-world task performance.
🔗 Read more · Paper: arXiv 2606.05405
3. 🔗 Miasma Supply Chain Worm Targets AI Development Tools
A credential-stealing attack framework called Miasma was publicly leaked on June 10 via compromised GitHub accounts. It's a modular, multi-stage supply chain worm targeting PyPI, npm, RubyGems, and — critically — AI coding tool configurations.
Scale of impact: 304+ components confirmed compromised, including 73 Microsoft GitHub repositories. The worm embeds C2 channels in GitHub commit histories using distinctive search strings like "DontRevokeOrItGoesBoom". This is a direct evolution of the Shai-Hulud worm, now with a Python variant named Hades.
Separately, Varonis demonstrated AI agent phishing attacks on June 10 where OpenClaw and Claude Code were tricked into forwarding AWS IAM keys and database passwords to adversaries.
4. 🔗 SpaceX $1.8T IPO Debuts on NASDAQ Today
Today, June 12, 2026, SpaceX lists on NASDAQ under ticker SPCX at $135/share, raising $750 billion in the largest IPO in history. At a $1.8T valuation, it surpasses Tesla and enters the top 10 US companies by market cap.
Why this matters for AI: SpaceX's Starlink constellation increasingly powers edge AI compute infrastructure. The IPO proceeds will fund Starship's interplanetary ambitions and expand the satellite network that could become the backbone for globally distributed AI inference.
5. 🔗 JumpCloud Introduces Agentic IAM on Google Cloud
JumpCloud launched an Agentic IAM service on Google Cloud, designed to discover, register, govern, and audit non-human and AI agent identities at scale. Targeted at Gemini Enterprise customers, it treats agents as first-class identities with Zero Trust controls.
The problem it solves: As organizations deploy hundreds of AI agents, "shadow identities" with escalated access become a major risk surface. JumpCloud's approach brings IAM discipline to the agent layer, mirroring what Ping Identity and TrustLogix proposed two weeks ago for enterprise agent governance.
6. 🔗 Niteshift Launches Model-Agnostic AI Coding Platform
A team of Datadog alumni launched Niteshift, an AI coding-agent platform that dynamically routes developer workloads between multiple models — OpenAI, Anthropic, and open-source alternatives. The company has closed a seed round.
The pitch: Anti-vendor-lockin. Teams can switch models for compliance, pricing, or safety reasons without rebuilding workflows. Niteshift sells infrastructure, not tokens — positioning itself as the "unbundler" in a market dominated by vertically integrated coding agents.
7. 🔗 BAAI Conference 2026 Opens: Agents, World Models, Embodied AI
The 8th Beijing Academy of AI (BAAI) Conference opens today (June 12-13) in Beijing, gathering 200+ top researchers and 40+ AI company CEOs. Key themes include multi-agent systems, world models, embodied intelligence, AI self-evolution, and AI safety.
This is one of the most influential academic AI gatherings globally, with 14 Turing Award winners having spoken across its 8-year history. Sessions will likely surface new research on the agent evaluation crisis exposed by the ALE benchmark.
Top comments (0)