The Intake
The Intake — Tuesday, July 8, 2026
On the substrate
Langflow lets authenticated users execute any other tenant's flows — confirmed active exploitation, CISA deadline July 10
CISA Known Exploited Vulnerabilities Catalog The Hacker News, July 8, 2026
If you're running Langflow in a multi-tenant setup, the July 8 CISA advisory named the isolation assumption that was wrong.
CVE-2026-55255 is an IDOR in Langflow's multi-tenant flow execution layer. An authenticated user can specify another tenant's flow ID in the API request and run that flow as their own. CISA added the vulnerability to its Known Exploited Vulnerabilities catalog on July 8; the federal remediation deadline is July 10. Active exploitation ran from June 22 to June 25. That activity has been attributed to a single threat actor at IP 45.207.216.55. That actor enumerated victim flows and extracted embedded LLM provider API keys and cloud credentials from other tenants' workflow definitions. The CVSS base score is 6.1. In observed attacks, the flaw chains with CVE-2026-33017, a separate unauthenticated remote code execution flaw in the same platform. That combination substantially raises the actual exposure. The Hacker News reports that no patched Langflow version is currently listed in the CISA KEV metadata.
If you're running Langflow with embedded API keys or cloud credentials in your workflow definitions, the pairing with the unauthenticated RCE flaw is the detail the 6.1 base score doesn't surface.
---
For operators
Artificial Analysis ranks GLM-5.2 top open-weights model; coding benchmark scores land within one point of closed frontier at one-sixth the price
Zhipu AI / Z.ai release announcement, June 13, 2026 The Decoder, June 17, 2026 Artificial Analysis Intelligence Index
If you build coding or long-horizon agentic tool-use pipelines against closed frontier models, Artificial Analysis's current Intelligence Index just moved GLM-5.2 to the top of the open-weights list.
GLM-5.2 is Zhipu AI's latest model, published June 13. The model carries an MIT license and a 1-million-token context window. Artificial Analysis's Intelligence Index — an independent third-party benchmarking service — ranks it as the current top open-weights model at 51 points. Zhipu's self-reported score on FrontierSWE — a software engineering benchmark — is 74.4. On the same benchmark, Zhipu reports Claude Opus 4.8 at 75.3. On SWE-bench Pro, Zhipu reports GLM-5.2 at 62.1. Zhipu's comparison puts GPT-5.5 at 58.6 on the same benchmark. Zhipu's reported MCP-Atlas tool-use score for GLM-5.2 is 77.0. Zhipu reports Opus 4.8 at 77.8 on that benchmark. The Decoder published an independent editorial evaluation on June 17. That evaluation confirms the FrontierSWE and Terminal-Bench 2.1 results. The Decoder also finds the model lags clearly behind Opus 4.8 and Gemini 3.1 Pro on Humanity's Last Exam — the gap is in general reasoning, not in coding or long-horizon tool-use. Pricing is approximately $1.67 per million input tokens; comparable closed models run $10 to $12.50 per million input tokens.
If your pipeline is specifically coding or long-horizon tool-use — not general reasoning — GLM-5.2 is the open-weights model now worth evaluating against your current closed-model allocation. The general reasoning tradeoff is where Humanity's Last Exam shows the gap.
---