Francis

There’s a lot happening in AI security right now, and most of it is being framed as “model security.” I think the bigger shift is this: AI defense is becoming a data control challenge first, and a model oversight challenge second. Here’s the context for why that matters: Enterprises are moving from “chatbots” to AI systems that touch real workflows. That means agents and copilots are increasingly reading, summarizing, transforming, and taking actions on top of: • regulated data (PII, PHI, PCI) • IP and deal data • source code and internal knowledge • customer communications and support systems

𝗜𝗱𝗲𝗻𝘁𝗶𝘁𝘆 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝘄𝗶𝗹𝗹 𝗰𝗵𝗮𝗻𝗴𝗲 𝗳𝗮𝘀𝘁 𝗶𝗻 𝟮𝟬𝟮𝟲. 𝗔𝗿𝗲 𝘆𝗼𝘂 𝗿𝗲𝗮𝗱𝘆? We are moving from identity governance built for slow human actions to a world where agents execute autonomous actions at machine speed. In our new research, we introduce Agentic Identity Access Platforms (AIAP): an end-to-end architecture that acts like a new SSO for agents, shifting governance from who logged in to why an action is happening, with task-scoped identities and permissions issued only when an authorized action is requested or in progress. We partnered with 5 vendors pushing this ecosystem forward: 1️⃣ @AstrixSecurity 2️⃣ @oasissec 3️⃣ @aembit_io 4️⃣ @TeamCyata 5️⃣ @silverfort Full report with case studies, implementation patterns, and our new ecosystem map: https://t.co/Vttya6Oquh

This past week, we’ve explored the identity spectrum for AI agents. Next week, our team is publishing one of our most comprehensive reports, covering the full spectrum and defining what it truly takes to secure agentic AI systems. Now we’re stepping back and addressing the bigger shift: AI security is no longer a feature. It requires a new architecture. Our very own Lawrence Pingree is leading this research, outlining what we call: 🔐 Unified Agentic Defense Platforms (UADP) These platforms converge: • Data Security (DSPM + DLP) • AI Security Posture Mgmt & Governance • Runtime prompt/endpoint agent protection • The full identity spectrum (human / non-human and agentic identities) • Intent-aware, JIT TRUST real-time enforcement This is about moving from fragmented point solutions to a unified control plane across endpoint, AI gateways, identity, and SOC. In this session, Lawrence Pingree will cover: ▪️ Why security teams are consolidating toward unified AI control planes ▪️ How UADPs manage human, machine, and autonomous agent identities together ▪️ What this shift means for architecture, governance, and compliance

Our firm predicts a major shift in how enterprises secure AI agents by H2 2026. Agent deployments are still early, but CISOs already see the next problem: vendor convergence and tool sprawl around a new actor class. What’s converging? 1/ Model security → expanding from LLM risk to agent operational risk (memory, tools, autonomy = lifecycle + identity). 2/ Identity and NHI → visibility into who an agent is, what it can do, and when access is revoked. 3/ Cloud and app security → agent-driven API execution + runtime behavior. 4/ Data security (DSPM/DLP) → stopping agents from exfiltration or misuse. 5/ Endpoint and browser → detecting execution edge abuse and enforcing guardrails. Bottom line: “AI security” is becoming agent lifecycle security, and agents are the new control plane. #AgenticAI #AISecurity #MCP #IdentitySecurity

We’re tracking a new identity security category that tackles an old gap: Identity Dark Matter. Think “EDR-style coverage” for identity apps. Most identity breaches don’t start at SSO. They start everywhere around it. Enterprises invested heavily in IAM: SSO/IdP, MFA, IGA, PAM. Policies pass audits. Attackers still move laterally and climb privileges. In our latest report by Lawrence Pingree, we define Identity Security Dark Matter as unmanaged identity artifacts and access paths that sit outside centralized IAM controls. Where it shows up: • Apps not fully onboarded to IGA • Fallback/local accounts that bypass MFA • Orphaned service accounts and API keys • Legacy protocols like NTLM • App-specific RBAC and hard-coded credentials • Partial SSO adoption and shadow identities Attackers live in the “unhappy paths” IAM doesn’t see, monitor, or govern. That creates a gap between intended policy and effective access. That gap is where lateral movement happens. We worked with @OrchidSecurity to go deep on this. Our take: Traditional IAM manages policy. We also need a control plane that brings visibility and control across identity items in apps, hosts, and runtime. It should tie identity to runtime signals, surface bypasses, and drive audit-ready fixes. If you’re a CISO or IAM leader and this feels familiar, you’re not alone. Full report here: https://t.co/AgSWOYf89i

In 2026, Cloud Security CNAPP and posture are dead. One vendor won that market. We need to focus on what is next. We believe the next big opportunity lies in cloud runtime / CDR. Most cloud security stacks still rely on snapshots: • CNAPP shows what was misconfigured • SIEM shows what happened • EDR shows what was executed But breakout times are now measured in minutes. Cloud environments change constantly. Stateless data from hours ago is no longer defensible. We’re seeing early adopters move toward a Cloud Twin architecture: a continuously updated, stateful model of identities, configurations, network reachability, and exposure in near real time. This shifts detection from log correlation to state-based reasoning and lets SOC teams prioritize alerts based on actual exploitability. Aqsa Taylor just published a case study on how @StreamSecurity applies this model to close the CDR gap without replacing CNAPP, SIEM, or EDR. Full report here: https://t.co/gIs9uS25hy

The biggest risks in AI model development! AI security isn’t just about securing the final product. The real threats happen during the development process. Three major risks are emerging: 1/ Data Poisoning—Attackers manipulating the learning process AI models rely on massive datasets—often from public sources. Adversaries inject poisoned data into datasets, forcing models to learn biased, unsafe, or manipulated behaviors. Once a model is poisoned, it’s nearly impossible to fix. Should enterprises start treating training datasets like a critical security asset? 2/ Backdoor Attacks—When AI models are compromised before deployment ; An attacker doesn’t need access to a deployed model to compromise it. They just need to modify it before deployment. Backdoors allow AI models to behave normally—until activated by a specific prompt. How can security teams detect backdoors in AI models before they go live? 3/ Model Drift—When AI becomes a security risk over time AI models learn from new data in production. That means a model can start secure but become a security risk over time. A model trained to be safe today could develop vulnerabilities just by interacting with adversarial users. How do security teams monitor for security drift in AI models? AI security isn’t a one-time check—it’s a constant validation process.

As enterprises race to deploy GenAI and LLMs, the risks around privacy, misuse, and model manipulation are no longer hypothetical—they're already here. This is why SACR is sharing our collaborative full market map of many AI security vendors building tools to secure enterprise AI across: 1️⃣ Governance Controls → Discover & protect sensitive data before AI models see it. 2️⃣ Model Security Controls → Catalog AI usage, scan models, and enforce policies. 3️⃣ AI Runtime Security → Monitor for prompt injection, API scraping, adversarial behavior—as it happens. The SACR Framework for Securing AI breaks the security stack into 3 actionable layers. We also note that a few vendors offer AI red teaming & penetration testing, but this is the core structure we see on the market. In general, this is the first time security is driving business velocity. AI security is no longer just operational—it’s a strategic imperative. Whether you’re a CISO, CTO, investor, or AI builder, this is your blueprint for navigating the AI security market.

The SOC of the future won’t be defined by detection engines. It will be defined by Data pipelines. What does this mean? Cribl is developing a new model for enterprise telemetry: a Telemetry Services Cloud designed to abstract the complexity of managing growing data volumes while giving teams control over how data is collected, processed, stored, and searched. The architecture is modular and vendor agnostic. Telemetry management decomposes into four key layers, each reflecting a broader shift across enterprise security and observability operations: 1/ Edge (collection) ➔ Move from centralized, high cost SIEM storage to open data lakes. 2/ Stream (processing and routing) ➔ Route only relevant telemetry to detection platforms. 3/ Lake/Lakehouse (storage) ➔ Normalize data with open schemas (e.g., OCSF). 4/ Search (query) ➔ Enable cross platform interoperability without re-ingesting data. Each module is independently deployable or can be combined into a unified pipeline. SOC's use this to filter low value events, enrich telemetry, and route data to specific tools. This reduces ingestion costs, improves data quality, and supports a policy driven approach to governance. IT operations, SRE, and platform engineering teams apply the same model to reduce noise, streamline routing, and implement tiered storage, with the option to store long tail logs for future analysis. Cribl does not position itself as a SIEM replacement. While its platform includes capabilities like flexible search and tiered storage, it lacks the advanced detection content and analytics typically found in SIEMs. Instead, it focuses on supporting those tools by improving data preparation and delivery. From a structural perspective, this model introduces a telemetry control plane that separates data infrastructure from application logic. It allows different teams to operate on the same data fabric while applying policies tailored to their roles, whether for cost control, compliance, or system performance.

Really great thread below (I particularly like the color on $NVDA vs $AMD) https://t.co/GBIi8613t1