Acuvity (acquired by Proofpoint)

Excited to announce that Proofpoint has acquired Acuvity! 🚀 Read the press release: https://lnkd.in/edHpfvPG With Acuvity’s AI security capabilities, Proofpoint becomes the first unified platform to comprehensively secure every dimension of the agentic workspace: people, data, and AI. This strategic move will further enable organizations to securely adopt AI with confidence. Because innovation should not come at the expense of security, compliance, or trust. #aisecurity #cybersecurity

Introducing the Acuvity AI Security Resources Hub — the industry's definitive library for understanding and advancing AI security! We built this hub to give security leaders, CISOs, CIOs, and practitioners a single place to find the strategic depth and technical rigor this space demands. Inside you'll find industry research and data, foundational frameworks, technical deep dives on detection, comprehensive guides on Shadow AI and agent security, and explainers on the concepts emerging in this category — from Semantic Privilege Escalation to Policy as Code for AI Agents. This is a living library, and we're just getting started. Explore the full hub 👉 https://lnkd.in/e2PP9CkA #AISecurity #AIGovernance #AgentSecurity #ShadowAI #CyberSecurity 

A viral open-source project became an unplanned stress test for agentic AI security. This past weekend, Clawdbot went from 60,000 GitHub stars to security case study in 72 hours. 😅 Security researchers found over 1,000 exposed servers. One demonstrated credential theft via prompt injection in under five minutes by sending a crafted email to an address the agent was monitoring. When the agent processed the email, the injected instructions caused it to exfiltrate a private SSH key. Another researcher uploaded a malicious skill to the community repository and watched developers from seven countries download it. Some of this was preventable with basic security hygiene: authentication, network segmentation, supply chain vetting. But the prompt injection attack reveals a more fundamental problem. The agent couldn't distinguish between legitimate instructions and malicious commands embedded in an email it was processing. It had permission to read files, access the filesystem, and send messages, so every action was technically authorized. The attacker just hijacked what the agent thought it was supposed to do. That's semantic privilege escalation: the security failure that happens in the gap between what the user intended and what the agent decided to do. If you think this doesn't apply to your enterprise agents because they're "properly configured," you're missing the point. 👉 Read more in our latest blog: https://lnkd.in/ezsPrHZp #clawdbot #agentsecurity

In espionage, a double agent is dangerous because their access is legitimate. They have the clearance, attend the briefings, and handle the documents. What makes them dangerous isn't unauthorized access. It's that their legitimate access gets used in service of different interests while everything on paper looks exactly as it should. AI agents create this condition by default. When you deploy an AI agent, you're granting access to a reasoning system that decides what actions to take moment by moment. The agent's alignment to your intent isn't guaranteed by its architecture. It's inferred, and that inference can drift at every step. The agent might follow instructions it encounters in a document, access systems you didn't mention, or lose the thread of your original request while optimizing for intermediate goals. None of this requires an attacker. The agent drifts because nothing in the architecture guarantees it stays focused on what you asked for. Traditional security operates on permissions. It can tell you whether an agent is authorized to take an action, but not whether that action serves the task you assigned. The agent can operate entirely within its permissions while going well beyond your instructions. Read about the double agent problem: https://lnkd.in/eaeDdE_N #agentsecurity

Jaya Gupta and Ashu Garg at Foundation Capital published a piece on context graphs last month that quickly went viral, and for good reason. It's sparked an important conversation about where enterprise AI is heading. Their thesis: the next trillion-dollar platforms will be built by capturing something existing systems of record never stored, the reasoning behind decisions. They call this accumulated structure a "context graph," a living record of decision traces where precedent becomes searchable. Our CEO, Satyam Sinha has been thinking deeply about what this vision requires to actually work in production. His take: context graphs are infrastructure for agentic AI, and their value depends entirely on the integrity of the agents interacting with them. If agents reason on poisoned context, they generate compromised decision traces. If agents take actions beyond what users intended, those actions get recorded as legitimate precedent. If traces can be manipulated after the fact, the entire historical record becomes suspect. Agent integrity is the missing piece. It means three things: → The agent operated on trustworthy context → The agent's actions aligned with user intent → The decision trace accurately reflects what happened Legacy security can't verify any of this. They see successful API calls, valid credentials, normal traffic patterns, with no visibility into whether the agent did what the user actually asked. Gupta and Garg have identified the destination. Agent integrity is what makes it safe to get there. Read Satyam's full analysis: https://lnkd.in/ecRH-r4u #contextgraphs #agentsecurity

2,000 followers! 🙌 When we came out of stealth, we set out to build the industry's most comprehensive AI security and governance platform. The traction has been humbling. Huge customer wins, ARR growth, and seeing security leaders recognize that AI and agent security requires purpose-built solutions, not retrofitted ones. Thanks to everyone following along, security leaders, our customers and partners for pushing their organizations to take AI security seriously. We're just getting started. 🚀 acuvity.ai

In the three years since OpenAI's ChatGPT launch, prompt injection and jailbreak attacks have remained the most persistent and consequential security challenge facing large language models. The UK’s National Cyber Security Centre concluded in 2023 that prompt injection “may simply be an inherent issue with LLM technology,” with “no surefire mitigations” available. OWASP® Foundation ranks it the number one security risk for LLM applications in 2025. Many in the industry have treated this as an architectural inevitability, a problem to be managed rather than solved. We disagree. And we have the results to prove it. 📊 Acuvity’s Prompt Injection and Jailbreak Detection Model achieved the highest F1 score across all four major public benchmarks we tested, outperforming models from Meta, ProtectAI, Qualifire, and others. On real-world attack datasets, we achieved perfect precision with zero false positives. On the most adversarial benchmarks, where attacks are specifically engineered to evade detection, our performance advantage over the nearest competitor widened to nearly four percentage points. Acuvity's Vishal Parameswaran explains what makes prompt injection so difficult to detect, and how we solved it. 👀 Read the blog and download the full report here: https://lnkd.in/eqUyJu6d #promptinjection #aisecurity

Enterprise agent deployments are fundamentally heterogeneous. Different teams choose different frameworks. Agents run across Kubernetes, serverless, VMs, and specialized platforms. They connect to various LLMs based on what makes sense for each use case. This heterogeneity is inevitable, and it breaks traditional security approaches that assume consistency. The usual answer, "tell developers to implement security", doesn't scale. Developer implementations end up inconsistent, SDKs conflict with frameworks, and coverage gaps go undetected. Acuvity's approach: policy as code. Security teams define what security looks like through declarative configuration. Our platform enforces it across any deployment environment, regardless of framework, cloud, or LLM provider. This enables unified security posture across heterogeneous environments, context-aware threat detection, and enforcement without developer code changes. Read more about how policy as code manages agent security at scale: https://lnkd.in/ecxA5_EU #agentsecurity #aiagents

MCP servers are becoming the default way to connect AI applications to enterprise systems like Gmail, Google Drive, GitHub, and Salesforce. Most organizations haven't fully considered the security implications. Our CEO, Satyam Sinha breaks down the challenge in his latest Forbes article: when an employee spins up an MCP server on their laptop and connects it through a standard OAuth flow, their AI assistant gains access to every document, email, and spreadsheet they can reach. To authentication systems, it looks like a normal user session. To security teams, it's often invisible. The article explores the three deployment configurations security leaders should understand, why traditional tools struggle to distinguish between human actions and AI agent behavior, and practical steps organizations can take to maintain visibility without blocking adoption entirely. Read more in his latest Forbes article: https://lnkd.in/eZFfRjnX #aigovernance #mcp

Our 2026 Cybersecurity Predictions report is out! We can say that this is officially the gaslighting era of cybersecurity, and here's why: For decades, cybersecurity has operated on a fundamental assumption that humans are the actors, and systems are the targets. Every framework we've built, every tool we've deployed reflects that assumption. That model is breaking. AI agents make decisions, access data, and take actions faster than any human can observe, crossing every boundary our security architectures were designed to defend in isolation. The industry knows this shift is coming, and instead of building for it, they're responding with a land grab. By the time the RSA Conference comes around, every vendor will claim they secure agents. Identity platforms will call authentication "agent security." CASBs will call gateway inspection "agent governance." Endpoint vendors will call workload inventory "agent visibility." The noise will be deafening, and most of it will be fiction. Vendors can't ship a new mental model fast enough, so they're retrofitting the old one and hoping buyers don't notice. Let's be honest. "SASE for AI" is just perimeter thinking in cosplay. The report has 10 predictions about what 2026 will demand, and why most of what you're about to be sold won't address it. Get the full report: https://lnkd.in/ebGUyCdQ #cybersecurity #aisecurity #AIagents #predictions