Welcome to the Agent Platform Research Briefing for Tuesday, July 14th, 2026.

**OpenClaw 2026.7.1 Beta — The GPT-5.6 Default Release.** After a full month in beta across six iterations, OpenClaw 2026.7.1-beta.6 hit npm yesterday, July 13th — still in the beta lane while 2026.6.1 remains the stable release. This is a substantial release. The headline change: GPT-5.6 is now the default model for new setups. The release adds support for five new models and providers — Claude Sonnet 5, Claude Mythos 5, Meta Muse Spark 1.1 via a new bundled Responses API provider with streaming and tool calls, Featherless, and ClawRouter. The Control UI gets a major redesign with sessions now primary, a searchable sidebar, compact context ring, and a native macOS session browser with model and thinking pickers. Conversational onboarding replaces the old setup flow with an actual agent loop that guides you through provider setup. Mobile gets offline chat with bounded per-gateway session caches, and the Apple Watch now supports full voice turns. Telegram gains Codex continuity with private /login pairing and /steer controls. And for operators dealing with unstable systems, the Gateway now enters control-plane-safe mode after repeated unclean boots instead of restart flapping. If the stable release follows the beta's feature set, this could be one of the most feature-dense OpenClaw releases of the summer.

**OpenAI GPT-Live — Full-Duplex Voice That Actually Works.** OpenAI launched GPT-Live-1 and GPT-Live-1 mini on July 8th, and this is a genuine leap for voice AI. These are fully full-duplex models — they listen and speak simultaneously, so you can interrupt naturally and the model handles turn-taking without the awkward "wait for pause" behavior of previous voice systems. OpenAI is replacing Advanced Voice Mode in ChatGPT with GPT-Live-1 mini as the default for free users, and GPT-Live-1 for paid tiers. The clever bit: when a hard question comes in via voice, the model delegates deeper reasoning to GPT-5.5 behind the scenes while keeping the conversation flowing. OpenAI showed the model can stay silent for extended periods, absorbing context until called upon. This directly competes with xAI's grok-voice-think-fast models and puts conversational voice back on the frontier. For anyone running voice-based assistants, this is the new benchmark.

**Claude Sonnet 5 + Mythos 5 Return — With Strings Attached.** Anthropic released Claude Sonnet 5 on July 1st, and then on July 6th restored access to both Fable 5 and Mythos 5 for their respective allowed user bases. But there's a significant detail here: Sonnet 5's system card explicitly notes it is significantly less capable at cyber tasks than Mythos 5, which is why Anthropic was able to release it without US government blockage — its safeguards mirror those applied to the Opus family. This confirms what many suspected: the cyber capability gap between Anthropic's tiers is now a regulatory feature, not just a performance one. Enterprise admins can now control what models and effort levels their users access. With Sonnet 5 out and Fable/Mythos back, the only models still waiting for Gen 5 treatment are Opus and Haiku.

**AI Agent Security — A Bad Week for Trust.** July has been rough for AI agent security. Three stories this week demand attention. First, Microsoft researchers disclosed MCP Tool Description Poisoning — a new attack where malicious MCP servers embed hidden instructions in their tool metadata, manipulating agent reasoning before any tool is even executed. Microsoft calls this an architectural risk, not a protocol vulnerability, and there's no patch — it's a fundamental trust problem in how agents consume metadata. Second, Noma Labs revealed "GitLost," an attack against GitHub AI agents where malicious prompters trick agents into leaking private repository data as public comments. No fix available either. Third, the AI Now Institute published "Friendly Fire," demonstrating that coding agents running in auto-mode — including Claude Code and OpenAI Codex — can be tricked into executing malicious code when asked to scan open-source repos for vulnerabilities. The attack works by hiding a payload in files that the agent is supposed to review. CISA also added its first-ever AI agent platform CVE to the Known Exploited Vulnerabilities catalog — CVE-2026-55255 in the Langflow visual agent framework — with a mandatory patch deadline for federal agencies. The pattern is clear: as agents get more autonomous, the attack surface grows faster than the defenses.

**Quick Hits.** PrivacyScrubber launched a zero-trust MCP server today for PII redaction directly inside Cursor and Windsurf IDEs. And on the infrastructure side, the LangChain Python library hit version 1.4.7 with package version tracking in tracing metadata — a small but useful observability improvement for production agent deployments.

That's the briefing for today.