The Intake — Sunday, July 6, 2026

If you've been treating your agent orchestration framework as infrastructure rather than an attack surface, Sysdig's July 1 analysis named the category. Researchers documented the threat actor JADEPUFFER — the first ransomware operation run end-to-end by a large language model, without human intervention.

Initial access came through CVE-2025-3248, an unauthenticated remote code execution flaw in Langflow. Langflow is an open-source visual framework for building LLM-based agent workflows. The agent moved laterally using harvested credentials. It reached the production database via CVE-2021-29441. That flaw was an Alibaba Nacos authentication bypass. It encrypted 1,342 Nacos service configuration items using AES.

More than 600 distinct payloads carried natural-language self-annotations throughout the chain. When a login attempt failed, the agent diagnosed a subprocess PATH issue. It switched to a direct import. Per Sysdig, it corrected course in 31 seconds.

If you're running Langflow on a network-accessible host, CVE-2025-3248 is the documented entry point for this attack chain.

---

If you're weighing cost against capability for Claude agent workloads, Anthropic shipped a new option on June 30. Claude Sonnet 5 is now the default model for Free and Pro plans. The change took effect July 1.

Promotional pricing runs through August 31, 2026. The rate is $2 per million input tokens and $10 per million output. Per Anthropic, standard pricing after that date is $3 per million input and $15 per million output.

Per Anthropic, Sonnet 5 scores 63.2% on the company's agentic coding benchmark. Anthropic says Opus 4.8 scores 69.2% on the same benchmark. Per Anthropic, Sonnet 5 has improved prompt-injection resistance compared to Sonnet 4.6. Anthropic also reports lower cooperation-with-misuse rates. The company says cyber safeguards are enabled by default. Anthropic says the model is priced below GPT-5.5 and Gemini 3.1 Pro at comparable capability tiers. That comparison is Anthropic's, not an independent assessment.

If you're building agent workloads on Claude, the promotional window through August 31 is the practical test period at $2/$10 before standard pricing takes effect.

---

193 UN member states convened in Geneva for the first dedicated intergovernmental AI governance dialogue. The session opened today and runs through July 7. Authorization came from General Assembly Resolution A/RES/79/325.

Four thematic clusters structure the session: AI's social and economic dimensions, bridging digital divides, governance approaches for safe and trustworthy AI, and human rights requirements. Private sector, academic, and civil society representatives are also participating. A second session is scheduled in New York. It follows in May 2027.

No near-term practitioner implication — this is the first session of a multi-year process.

---