DeepKeep

DeepKeep now integrates with LangChain. LangChain developers can add DeepKeep’s AI Firewall as agentic guardrails using the same middleware model that controls their agent workflows. The integration gives teams a direct path from agent development to production security. Less custom security plumbing and more control where the agent runs. More info here: https://lnkd.in/d6GNfZX3 #AISecurity #AgentSecurity #LangChain

The AI agent stampede reaches production. #AISecurity #AIAgents

AI firewall policy needs more than “detect and block.” Once risk is detected, the response should fit the context: 🔹Block 🔹Redact 🔹Replace 🔹Alert Same sensitive value. Different path. Different action. Mask it from the model, preserve it for the authorized user, block it from leaving through the wrong tool. Simple breakdown here: https://lnkd.in/dsy-a3fV #AISecurity #AIFirewall

Europe's Multilingual Reality Exposes AI Security Gaps: https://bit.ly/4vNHblA by Alexander Culafi

Your AI may support 100+ languages. That doesn't mean your AI security layer can enforce policy across them. For global teams, language isn't an edge case. AI workflows process support tickets, contracts, invoices, HR records, screenshots, regional knowledge bases, and customer messages in whatever language the business runs on. The issue is that attack intent can survive language shifts better than security controls do. A prompt injection can be softened by translation, hidden in retrieved content, or carried through a document before an agent treats it as context. The question is not whether the model can answer in Japanese, German, Spanish, or Korean. It's whether the security layer reaches the same enforcement decision when the same attack intent appears across languages, formats, and workflows. That is what multilingual AI security has to prove: prompt injection detection, PII protection, unsafe output handling, and tool-use enforcement at runtime, before the model acts or the agent calls the tool. Language support is a product feature. Multilingual enforcement is a security requirement. https://lnkd.in/dJrJPzMR #AISecurity

AI security architecture is not a sticker you slap between the app and the model. Inline enforcement is useful when something needs to be stopped before it reaches a model, tool, or workflow. It is also where latency, fail-open choices, policy mistakes, and inspection costs become very real. Out-of-band control can enforce in real time too. It just depends on what's actually routed through it, and what isn't. The question is not "inline or out-of-band?" - it's where prevention is mandatory, where visibility is enough, and where the diagram is doing more work than the control. https://lnkd.in/dxNEeVq8 #AISecurity #AIGovernance #AgenticAI

Q3 security roadmap: ask nicely. Q4: ask nicely, but in bold. #AIGovernance

The EU AI Act turns AI security into an evidence problem. For high-risk AI systems, organizations need to show how risk is managed, how behavior is monitored, how records are kept, and how controls operate in practice. That is not only legal work. It creates a technical burden for security teams. Can you show what was tested? What was logged? Which policy was violated? What was blocked, redacted, replaced, or alerted on? Which model, application, agent, or employee AI use path was involved? If legal, audit, or leadership asks how your AI systems behaved, the answer can't be "we'll check." It has to be evidence, on demand. #AISecurity #EUAIAct #AIGovernance

Three weeks until Black Hat USA. By then, plenty of AI agents will have been deployed, connected to tools, and trusted with actions nobody fully mapped. Efficient. We’ll be at Booth 5706, showing what enterprise AI security looks like when it covers the full lifecycle, not just the prompt box. And yes, we’ll be putting AI systems under pressure with Vibe AI Red Teaming, because the most interesting vulnerabilities rarely appear when everyone follows the happy path. Las Vegas, we’re getting closer. Black Hat USA | Booth 5706 #AISecurity #BlackHat

SIEM had this problem. So did early EDR. Alert volume outstripped what a security team could triage, and vendors who never solved it lost customers at renewal. Gartner's own warning, in their latest Emerging Tech report on agent and LLM security (by David Senf and Mark Wah), is that agent security is walking into the same trap. Just faster, and with fewer people equipped to catch it. The market growing around this is real. Securing AI spend is on a path to nearly 6x growth by 2030, fast enough to pass the data security market by 2029 and application security by 2030. But most of that spend goes toward vendors extending yesterday's tools, IAM, firewalls, DLP, to cover AI, not toward anyone built for how agents actually operate. Catching a bad prompt is table stakes. Gartner is direct about what's needed beyond it: security controls have to go past standard text filtering to catch manipulation at the decision layer, and the stakes for prompt injection and goal hijacking are far higher once an agent can act on it. Multimodality, agent behavior analysis, fine-grained authorization, and context-aware enforcement are the gaps they call out as underserved, and the ones about to matter most. An agent's real risk doesn't live only in the prompt. It lives in what it does next. #AISecurity #AgenticAI #LLMSecurity