โ† Back to all episodes
Agent Platform Research โ€” February 25, 2026
February 25, 2026 ยท ๐Ÿ”ฌ Research

# Agent Platform Research Briefing โ€” 2026-02-25

## New Stories

---

### 1. OpenClaw 2026.2.23 โ€” Security Hardening Patches Endor Labs CVEs, Adds Claude Opus 4.6

Status: NEW

OpenClaw released v2026.2.23 yesterday, tagged by steipete. Critically, this release **patches the six CVEs** reported by Endor Labs on Feb 23 (CVE-2026-26322 SSRF/7.6, CVE-2026-26319 webhook auth/7.5, CVE-2026-26329 path traversal) โ€” all previously unpatched. Key security fixes:

- **SSRF policy** now defaults to `trusted-network` mode (breaking change; run `openclaw doctor --fix` if on private network)

- HTTP `Strict-Transport-Security` headers added for direct HTTPS deployments

- Config snapshots redact `env.*` and `skills.env.*` keys โ€” prevents API key leaks in logs and diagnostics

- OTEL telemetry scrubs API keys before export

- Skills packaging rejects symlink escapes and XSS-vulnerable gallery prompts

- Command approval required for obfuscated injection patterns

- Gateway pairing: operator.admin tokens now satisfy operator.write requests (fixes legacy device scope-upgrade loops from 2026.2.19)

- Auto-abort stop phrases expanded (STOP OPENCLAW!!!, please stop, stop agent, etc.)

- Sandbox media rejects hard-linked tmp aliases including symlink-to-hardlink chains

AI capability additions:

- Claude Opus 4.6 support via Kilo Gateway (full auth stack)

- Moonshot Kimi native video analysis + enhanced citation extraction

- Vercel AI shorthand Claude reference normalization

**Bottom line:** Update is now urgent for security. The three high-severity Endor Labs CVEs have confirmed patches in 2026.2.23.

---

### 2. Anthropic Claude Cowork โ€” Department-Level Enterprise Plugin Expansion

Status: NEW

Announced Tuesday Feb 24, rolling out now. Anthropic expanded Claude Cowork with a full suite of domain-specific plugins and enterprise connectors. Previously the platform had a general-purpose work assistant; now organizations can deploy purpose-built agents for:

- **Finance / Investment Banking** โ€” financial analysis workflows

- **Legal** โ€” document review, contract analysis (LegalZoom integration)

- **HR** โ€” onboarding, policy Q&A

- **Engineering** โ€” code review, technical documentation

**Enterprise connectors:** Google Drive, Gmail, DocuSign, FactSet, LegalZoom. Admins can deploy and manage plugins centrally.

**Market reaction:** Software and cybersecurity stocks dropped sharply on the announcement โ€” notably ThomsonReuters and LegalZoom โ€” as the market interprets this as direct competition with vertical SaaS. Anthropic framing: "enables solo founders and small teams to use Claude like a virtual workforce."

**Significance:** This is Anthropic's most aggressive enterprise push yet, moving beyond general assistant into workflow automation territory that directly competes with established SaaS.

---

### 3. IBM + Deepgram โ€” Real-Time Voice Lands in watsonx Orchestrate Agent Builder

Status: NEW โ€” Voice AI

IBM and Deepgram announced a partnership February 24, making Deepgram IBM's first voice technology partner. Deepgram's real-time speech-to-text and text-to-speech will be embedded directly into watsonx Orchestrate's Agent Builder, enabling enterprise customers to build voice-driven agentic workflows without assembling a custom speech pipeline.

**IBM shares climbed 5%** on the announcement. Deepgram gains enterprise distribution; IBM fills a gap in its agentic platform (Orchestrate was text-only for agent-to-human interactions).

**Why it matters for agent builders:** Enterprise voice agents have historically required stitching together separate ASR, LLM, and TTS services. Deepgram-in-Orchestrate makes voice a first-class orchestration primitive alongside tool calling and memory โ€” a pattern the rest of the market will likely follow.

---

## Status Updates on Active Stories

- **DeepSeek V4 Lite + GPT-5.3 "Garlic" (Feb 26 targets):** DeepSeek's mid-February launch window passed without release. No new date announced as of this morning. The Feb 26 rumored date appears to have also slipped. Still watching.

- **OpenClaw Endor Labs CVEs:** Patched in 2026.2.23 (see Story 1 above). Updating status to `resolved`.

- **MIT Agentic AI Safety Study:** No policy response or vendor follow-up yet.

---

## No New Developments

- **LangChain / LangGraph / CrewAI / AutoGen / DSPy:** No notable releases in the past 24-48h. Coverage is evergreen overview articles, not news.

- **MCP ecosystem:** CIO.com notes MCP is "on every executive agenda" ahead of RSA 2026, but no new servers or integrations of note. SDK CVE fix (v1.26.0) still being absorbed.

- **NVIDIA PersonaPlex-7B:** No new benchmarks or adoption news. Skip per standing note.

- **MiniMax Speech 2.6:** No update, per hold period.

- **OpenAI Realtime API:** New voices spotted in developer community forums (names not yet documented), but no formal announcement.