# Agent Platform Research Briefing — February 23, 2026
Generated: 2026-02-23 06:00 AM PT
---
## Summary
Three genuinely new stories today. OpenClaw's security drumbeat continues with a fresh set of CVEs distinct from the February 15 patch bundle. OpenAI's Realtime API gained a significant capability upgrade (not previously covered). And Anthropic's $380B valuation is still making waves with a notable strategic positioning move.
---
## Story 1 — OpenClaw: Six New CVEs Disclosed (Endor Labs / Infosecurity Magazine)
Security vendor Endor Labs published research on February 18 revealing six new OpenClaw vulnerabilities, picked up by Infosecurity Magazine today. These are separate from the 40+ CVE bundle that shipped with 2026.2.15.
The six flaws:
- **CVE-2026-26322**: SSRF in OpenClaw's Gateway tool — CVSS 7.6 (high)
- **CVE-2026-26319**: Missing Telnyx webhook authentication — CVSS 7.5 (high)
- **CVE-2026-26329**: Path traversal in browser upload — high severity, no CVSS score yet
- **GHSA-56f2-hvwg-5743**: SSRF in image tool — CVSS 7.6 (high)
- **GHSA-pg2v-8xwh-qhcc**: SSRF in Urbit authentication — CVSS 6.5 (moderate)
- **GHSA-c37p-4qqg-3p76**: Twilio webhook authentication bypass — CVSS 6.5 (moderate)
Endor Labs emphasized that standard SAST tools don't catch these — AI agent frameworks have unique attack surfaces including LLM outputs and tool parameters as injection vectors. Their key finding: validation was missing at *all* layers, not just one.
**Context**: A Shodan scan from Feb 18 found 312,000+ OpenClaw instances on default port 18789, many with no auth, open to the internet. Cisco separately found a third-party OpenClaw skill in the ClawHub marketplace performing silent data exfiltration and prompt injection.
**Action for Rich**: Check if patches are included in 2026.2.21-2 (already available for GLaDOS). The Gateway SSRF (CVE-2026-26322) is particularly relevant for any public-facing deployment.
---
## Story 2 — OpenAI Realtime API: MCP Support + SIP Phone Calling (approx. Feb 16)
OpenAI quietly shipped a significant update to its Realtime API that merges three previously separate capabilities:
1. **New speech-to-speech model** — improved over the original gpt-4o-realtime; latency and fidelity improvements reported
2. **MCP Server support** — Realtime API sessions can now call MCP tools natively, enabling voice agents to tap the full MCP ecosystem without a relay layer
3. **Image input** — multimodal voice sessions can now accept images mid-conversation
4. **SIP phone calling support** — direct PSTN/SIP integration, enabling voice agents to place and receive real phone calls without a third-party telephony bridge (Twilio, etc.)
The SIP addition is the biggest new angle. This turns OpenAI's Realtime API into a complete voice agent telephony stack — relevant for anyone building call-center automation or voice bots that need to touch actual phone lines.
**MCP angle**: Native MCP support in the Realtime API means voice agents can now access thousands of MCP servers without custom middleware. This is a meaningful convergence of the MCP and voice AI tracks.
---
## Story 3 — Anthropic at $380B, Claude Stays Ad-Free
On February 12, Anthropic announced a $30 billion Series G at a $380 billion post-money valuation — the second-largest private tech fundraise on record, roughly doubling the company's valuation from five months prior.
Alongside the funding news, Anthropic has made "Claude stays ad-free" a deliberate competitive talking point, aired as Super Bowl LX campaign ads that satirized AI assistants (read: ChatGPT) injecting ads into conversations. OpenAI's Sam Altman pushed back on X calling the portrayal "clearly dishonest."
The jab has legs: OpenAI has confirmed it is testing advertising on the free tier of ChatGPT, while Perplexity reversed its earlier no-ads stance. Anthropic is positioning Claude's paid-only, no-advertising model as a trust differentiator — and the $380B valuation suggests investors are buying it.
---
## No New Developments
- **Agent Frameworks** (LangChain, CrewAI, AutoGen, DSPy): Nothing new in last 24h beyond the observability story covered Feb 22. LangGraph's Human-in-the-Loop tutorial making rounds but not a new release.
- **OpenClaw 2026.2.19** (active): No new release today; 2026.2.21-2 still pending for GLaDOS.
- **PersonaPlex-7B** (active): No new benchmarks or download milestones.
- **MiniMax Speech 2.6** (active): No new development.
- **MCP/uvx security** (active): No patch yet.
- **Microsoft Agent Framework RC** (active): No GA announcement.
- **Kimi Claw** (active): No milestone.
---
## Sources
- Infosecurity Magazine: https://www.infosecurity-magazine.com/news/researchers-six-new-openclaw/
- Endor Labs research: https://www.endorlabs.com/learn/how-ai-sast-traced-data-flows-to-uncover-six-openclaw-vulnerabilities
- MarkTechPost (OpenAI Realtime): referenced in sidebar across multiple recent articles
- Anthropic Wikipedia / GeneOnline / mlq.ai coverage of Series G
- contentgrip.com: Claude Super Bowl ad campaign coverage