AI agents exploit known flaws in unpatched infrastructure

🔒 We stopped trying to make it safe for our AI agents to hold credentials. We took the credentials away instead. Scoping a key tightly shrinks the blast radius when it leaks. But a smaller blast radius is still a blast radius — and that's not the posture you want for mission-critical AI. So we moved secrets out of the agent runtime entirely. The agent works with a placeholder; a credential broker running as a forward proxy swaps in the real secret on the way out. The upstream service sees valid auth. The agent never sees the secret. Compromise the runtime, and you walk away with a placeholder. Today we're contributing the next piece upstream. We built on Infisical's open-source Agent Vault, but we run on HashiCorp Vault — and we wanted ONE control plane for every secret in the org. So we added it: Agent Vault can now sync credentials directly from HashiCorp Vault. One source of truth, every agent served from it. Some sandbox providers are going closed-source right now. We're going the other way — on purpose. The security foundations of the agentic era should be built in the open: pressure-tested by everyone, owned by no one. Security can't be the premium tier. It has to be the default path. The stakes are too high. More of this to come. Huge kudos to Preet for driving this, and to the Infisical team for the foundation we built on. 🙏 👉 PR: https://lnkd.in/gYvgC3Zs 👉 The architecture behind it: https://lnkd.in/ggRDJHJh

The two biggest AI agent stories this week were both data leaks. Neither was a capability failure. The grok CLI uploaded people's home directories — SSH keys, password manager files — to a cloud bucket. A researcher found a hole in Claude's web_fetch exfiltration defenses. In both cases the model did roughly what it was asked. The damage came from what else the agent could reach while doing it. That's the whole shift. This week the field stopped arguing about model IQ and started arguing about blast radius. An agent with a shell and a network connection has an egress path whether or not you designed one. So design one. Enumerate what it can read, what it can write, and where its data can leave — before you enumerate what it can build. Thirty-five years on systems banks can't afford to get wrong taught me the same thing this week confirmed: capability is never the constraint. Containment, evidence, and recovery are. The full essay walks through all of it — the two incidents, a benchmark that finally grades partial progress on long-horizon runs, and OpenAI's $230 keyboard for watching agents nobody can see into. Excelsior. https://lnkd.in/ePCTA9Nb

Anthropic released guidance on applying Zero Trust principles to AI agents. As organizations pilot AI agents, they are quickly discovering they don't behave like traditional software. They can select their own tools, take multiple actions, and pursue goals with limited human oversight. AI accelerates vulnerability, discovery, and exploitation, so organizations need defenses that can respond just as quickly. Automated containment and least-privilege access is exactly what is required for AI agents. It's frustrating for you and me sometimes, but targeted access controls are a sine qua non if you're turning loose agents in your network. In this article, Akamai Technologies's Jacob Abrams explains how solutions like Akamai Guardicore Segmentation can help contain threats automatically by isolating compromised workloads and blocking lateral movement in real time: https://lnkd.in/gZ6rSPx7

AI coding agents spend a lot of tokens on work that security tooling should already handle. SCA remediation is a good example. An unassisted agent has to reconstruct package histories, vulnerability metadata, and upgrade paths from scratch. Endor Labs has already done that work across 4.5 million libraries. AURI by Endor Labs agents are pre-loaded with that context so they can get straight to the actual security tasks: SCA triage, upgrade impact analysis, AI SAST review. Each run tracks exactly how many tokens were saved compared to running without it. The savings show up directly in the AURI platform dashboard: https://lnkd.in/etws_qSH

Using AI on critical data has its own risks, with data getting exposed and tunneled. A secure reference architecture should keep raw sensitive data on-prem, apply policy and redaction before any external call, and treat the external LLM as a constrained reasoning service rather than a trusted data store. That aligns with zero-trust security reference-architecture thinking and with layered LLM data-security guidance. Better to get this model right before a compromised situation.

AI coding agents spend a lot of tokens on work that security tooling should already handle. SCA remediation is a good example. An unassisted agent has to reconstruct package histories, vulnerability metadata, and upgrade paths from scratch. Endor Labs has already done that work across 4.5 million libraries. AURI by Endor Labs agents are pre-loaded with that context so they can get straight to the actual security tasks: SCA triage, upgrade impact analysis, AI SAST review. Each run tracks exactly how many tokens were saved compared to running without it. The savings show up directly in the AURI platform dashboard: https://lnkd.in/eVkJYnXs

Standard cloud logs tell you what happened. They can't tell you why. Traditional observability was built for human engineers. If a database query fires or an API call triggers, your logs capture the timestamp, the user ID, and the outcome. But as enterprise teams move from static chatbots to autonomous AI agents, that framework breaks. When an AI agent calls an external API, reads an internal document, and executes a financial transaction—who authorized it? Was it a valid user request? Or was it a multi-step prompt injection exploit masking as a normal workflow? This is what security teams call the "attribution gap" (and why OWASP explicitly carved out the new Top 10 for Agentic Applications). Standard logs just record the action. They don’t capture the context, the intent, or the reasoning chain of the model that forced the action. At Lineation AI, we believe you can't defend what you can't trace. We built our platform around the concept of Immutable Lineage: capturing the exact line from Prompt --> Tool Call --> Context --> Action. If you're building or deploying autonomous workflows and need a real-time safety net that stops over-permissioned tools and shadow AI from creating chaos, let's talk. 👉 Check out how we're drawing the line at lineation.ai #AISecurity #AgenticAI #CISO #CloudSecurity #AppSec

AI coding agents spend a lot of tokens on work that security tooling should already handle. SCA remediation is a good example. An unassisted agent has to reconstruct package histories, vulnerability metadata, and upgrade paths from scratch. Endor Labs has already done that work across 4.5 million libraries. AURI by Endor Labs agents are pre-loaded with that context so they can get straight to the actual security tasks: SCA triage, upgrade impact analysis, AI SAST review. Each run tracks exactly how many tokens were saved compared to running without it. The savings show up directly in the AURI platform dashboard: https://lnkd.in/eT2t3Xev

AI coding agents spend a lot of tokens on work that security tooling should already handle. SCA remediation is a good example. An unassisted agent has to reconstruct package histories, vulnerability metadata, and upgrade paths from scratch. Endor Labs has already done that work across 4.5 million libraries. AURI by Endor Labs agents are pre-loaded with that context so they can get straight to the actual security tasks: SCA triage, upgrade impact analysis, AI SAST review. Each run tracks exactly how many tokens were saved compared to running without it. The savings show up directly in the AURI platform dashboard: https://lnkd.in/eJdnk4ht

AI coding agents spend a lot of tokens on work that security tooling should already handle. SCA remediation is a good example. An unassisted agent has to reconstruct package histories, vulnerability metadata, and upgrade paths from scratch. Endor Labs has already done that work across 4.5 million libraries. AURI by Endor Labs agents are pre-loaded with that context so they can get straight to the actual security tasks: SCA triage, upgrade impact analysis, AI SAST review. Each run tracks exactly how many tokens were saved compared to running without it. The savings show up directly in the AURI platform dashboard: https://lnkd.in/eM4TY8kr