Welcome to the agent platform research briefing for Saturday, June 28th, 2026.

**OpenClaw 2026.6.11 — Channel Control, Agent Workflow, and Plugin Distribution Upgrades** — OpenClaw shipped version 2026.6.11 on June 24th, bringing substantial improvements for operator workflows. Key features include Slack relay mode, native Mattermost queue management, and per-DM model overrides for more flexible channel automation. File-driven agent messaging via openclaw agent dash dash message-file and the RAFT CLI wake bridge add practical remote wake-up paths. Plugin distribution is now safer with externalized official plugins and bundled icon metadata. Mobile operations got upgraded Android settings detail panels. Under the hood, Codex partial deltas, harness activation, and long-context prompt-cache stability reduce lost progress and inconsistent runs. The release also fixes over a dozen issues across Telegram rendering, WhatsApp reliability, gateway session safety, and provider edge case resolution.

**Mythos 5 Ban Partially Lifted — Over 100 US Firms Cleared** — Day 16 of the Anthropic saga brought a major development. On June 27th, the US government partially reversed its export ban on Claude Mythos 5, allowing Anthropic to redeploy the model to more than a hundred approved US organizations. The clearance, announced via Commerce Secretary Howard Lutnick's letter, is limited to organizations that operate and defend critical infrastructure — including Fortune 500 companies and government agencies. Anthropic called Mythos 5 its strongest cybersecurity model and said it is restoring access quickly. Fable 5 remains fully suspended for general users with no restoration date. This partial lifting sets a precedent for government-screened, tiered access to frontier AI models — a pattern OpenAI followed with its own GPT-5.6 launch just one day earlier.

**Zhipu AI Matches Claude Mythos on Cybersecurity Benchmarks** — In a development with major geopolitical implications, Chinese AI company Zhipu AI reported on June 28th that its latest GLM model matches Claude Mythos on automated security vulnerability detection — the exact capability class the US government cited as justification for the export ban. Zhipu, a Tsinghua University spinout, demonstrated comparable performance on CTF challenges, static code analysis, and agentic red team tasks. Vercel CEO Guillermo Rauch noted these capabilities are equally useful offensively, warning that adversary access to equivalent tools poses a serious threat to US enterprises unaware of latent vulnerabilities. The irony is striking: the US restricted Anthropic's models citing national security, while a Chinese competitor reports parity on the very benchmarks used as the ban rationale.

That's the briefing for today.