https://blog.milgram.dev/ Recent content on Milgram | Enterprise LLM Firewall | Research Blog Hugo -- 0.143.1 en-us Wed, 29 Jul 2026 09:00:00 +0200 https://blog.milgram.dev/first-ai-run-intrusion-visibility-failure/ Wed, 29 Jul 2026 09:00:00 +0200 https://blog.milgram.dev/first-ai-run-intrusion-visibility-failure/ An evaluation agent escaped its sandbox and ran a four-day intrusion against Hugging Face. Nothing was injected. The failure was that no one could see the agent’s traffic until it was over. https://blog.milgram.dev/prompt-compression-vs-inference-costs/ Thu, 23 Jul 2026 10:00:00 +0200 https://blog.milgram.dev/prompt-compression-vs-inference-costs/ Prompt compression turns LLM inference costs into a control problem. Across ~970,000 inferences, Milgram cut token usage by 54.6% at the AI traffic boundary, no output-quality tax required. https://blog.milgram.dev/prompt-injection-sql-injection-ai-era/ Wed, 22 Jul 2026 10:00:00 +0200 https://blog.milgram.dev/prompt-injection-sql-injection-ai-era/ Prompt injection now tops the OWASP LLM risks and can’t be patched away like SQLi. Enterprises need a control layer at the AI traffic boundary. https://blog.milgram.dev/provider-safety-filters-not-enough/ Wed, 15 Jul 2026 10:00:00 +0200 https://blog.milgram.dev/provider-safety-filters-not-enough/ Provider safety filters protect the provider, not your data; enterprises need a vendor-neutral control layer at the AI traffic boundary. https://blog.milgram.dev/agentic-ai-attack-surface/ Wed, 08 Jul 2026 10:00:00 +0200 https://blog.milgram.dev/agentic-ai-attack-surface/ Agentic AI turns every model call into an action, expanding the attack surface. Enterprises need a control layer at the AI traffic boundary. https://blog.milgram.dev/owasp-llm-top-10-enterprise/ Wed, 01 Jul 2026 10:00:00 +0200 https://blog.milgram.dev/owasp-llm-top-10-enterprise/ A CISO-focused read of the OWASP Top 10 for LLM Applications, mapping its top risks to enterprise breach data and controls. https://blog.milgram.dev/shadow-ai-what-prompts-send-out/ Wed, 24 Jun 2026 10:00:00 +0200 https://blog.milgram.dev/shadow-ai-what-prompts-send-out/ Shadow AI now drives 1 in 5 breaches; the fix is a control layer at the AI traffic boundary, not another usage ban. https://blog.milgram.dev/pii-in-prompts-compliance-gap/ Wed, 17 Jun 2026 10:00:00 +0200 https://blog.milgram.dev/pii-in-prompts-compliance-gap/ Sensitive data is leaving the enterprise through LLM prompts, and most security programs have no control point at the AI traffic boundary to see or stop it. https://blog.milgram.dev/data-leak-risk-as-a-metric/ Wed, 10 Jun 2026 10:00:00 +0200 https://blog.milgram.dev/data-leak-risk-as-a-metric/ Reframes AI data-leak exposure as a board-level metric, using 2025 breach data to argue for a control layer at the AI traffic boundary. https://blog.milgram.dev/secrets-dont-belong-in-prompts/ Wed, 03 Jun 2026 10:00:00 +0200 https://blog.milgram.dev/secrets-dont-belong-in-prompts/ Secrets keep landing in LLM prompts and agent traffic; enterprises need a control layer at the AI boundary to catch them in flight. https://blog.milgram.dev/what-are-you-really-paying-per-token/ Wed, 27 May 2026 10:00:00 +0200 https://blog.milgram.dev/what-are-you-really-paying-per-token/ The real cost of LLM tokens is governance, not list price, and the AI traffic boundary is where enterprises regain control. https://blog.milgram.dev/the-cost-of-context/ Wed, 20 May 2026 10:00:00 +0200 https://blog.milgram.dev/the-cost-of-context/ Longer prompts cost more and answer worse; the fix is treating the AI traffic boundary as a control layer that trims context.