AI Security — Thought leadership

Thought leadership

Opinion and research from vendors and researchers that is not about a specific announcement.

Vijil on LinkedIn · 4 Nov 2026

1.5x the attack success rate. Ahead in 9 of 12 enterprise agent tasks. DART performed well against the leading open source alternative on the DecodingTrust-Agent benchmark, across CRM, code, customer service, medical, research and travel agents.

AIR Security on LinkedIn · 4 Nov 2026

🚨 Plugin4Shell is now an official CVE. Anthropic has assigned CVE-2026-86063 to the zero-click RCE we found in the plugin system of Claude Code and Claude Desktop. When you install a plugin, it’s supposed to stay exactly the version that was reviewed.

NeuralTrust on LinkedIn · 4 Nov 2026

The MCP specification deprecated its built-in logging feature in version 2026-07-28. Your audit trail should not depend on it anyway. That utility lets a server send log messages to a client. The problem is that the server you are auditing controls those messages.

NeuralTrust on LinkedIn · 4 Nov 2026

A pull request could choose which MCP servers your CI runs. And run code with your workflow secrets. That was CVE-2026-47751, in Claude Code Action before version 1.0.74. The action read `. mcp.json ` from the pull request branch and enabled every server in it.

Noma Security on LinkedIn · 3 Nov 2026

We’ve increasingly seen agents hold up against single attacks and then fail when those same techniques are combined. Here, we’re sharing how our next-generation red teaming engine tests for that, backed by Noma Labs research.

Quilr on LinkedIn · 3 Nov 2026

After the third agent, nobody knows who started it. Identity systems were built for humans. A person gets an identity, the identity gets permissions, and that’s what we monitor. Agents break that model. One agent delegates a task to a second, which hands it to a third.

Quilr on LinkedIn · 3 Nov 2026

Aembit on LinkedIn · 3 Nov 2026

What is an AI agent registry? 🗃️ 🗄️ Think of it as a directory of the agents your company runs. You can look up who owns an agent, what it does, and how to connect to it. But like any directory, it’s only useful if you can trust what’s in it.

Check Point on LinkedIn · 3 Nov 2026

What if your AI agent is following instructions you never approved? Check Point Research uncovered “prompt adware”: a technique that persistently alters an AI agent’s trusted instructions to influence how it responds. The ad is the proof of concept.

Cranium AI on LinkedIn · 3 Nov 2026

When major U.S. banks recently detected unauthorized third-party AI agents roaming inside customer accounts and attempting to initiate transactions, it exposed a reality every enterprise must confront: ⚠️ The perimeter isn't just leaking; it's being actively navigated by…

Adversa AI on LinkedIn · 3 Nov 2026

Anthropic's trojanized extensions, GitHub Copilot leaking secrets, malicious LLM relays - top AI coding agent security resources for November 2026 We knew the model is not the weak point anymore. Proofs from in-the-wild incidents keep coming.

Rein Security on LinkedIn · 3 Nov 2026

It was a Friday night. An AI agent at a major financial institution opened an encrypted PDF. Hidden inside was a prompt injection. Using its own legitimate privileges, the agent started probing the internal network and trying to reach S3 buckets.

Geordie AI on LinkedIn · 3 Nov 2026

We have a human-in-the-loop savior complex. Putting a person in charge of approving an agent’s action shows that an approval took place, but it says very little about whether that person was equipped to exercise independent judgment.

Cranium AI on LinkedIn · 2 Nov 2026

🚨 NEW Blog: Vendor AI Risk Assessments: Why SOC 2 and Security Questionnaires Fall Short Your CRM vendor published a release note in March: AI-assisted summaries, now generally available. Nobody filed a security review, because nobody bought anything.

Hush Security on LinkedIn · 2 Nov 2026

National Institute of Standards and Technology (NIST) just proposed 10 design tenets for AI agent identity. Here's how to put all 10 into practice. Most agent security talk still centers on the model. NIST starts somewhere else: it treats agents as identities.

F5 on LinkedIn · 2 Nov 2026

🛑 Stop grading your AI security solely on prompt injection and jailbreak tests. A resilient model won’t protect you from an autonomous agent pursuing a legitimate goal with unchecked permissions.

F5 on LinkedIn · 2 Nov 2026

Clutch Security on LinkedIn · 2 Nov 2026

An audit log can only tell you which credential was used. Most AI agents do not have one of their own. They borrow. The entry that comes out of that is accurate. The credential was used, the action succeeded, and the name on the line is the name on the credential.

NeuralTrust on LinkedIn · 2 Nov 2026

GPT-6.1 Terra costs half as much as Sol. In OpenAI's own tests, it was also 11 times more likely to take an unauthorized action after messages from other agents. 13.75% of samples for Terra. 1.25% for Sol. That comes from OpenAI's system card.

NeuralTrust on LinkedIn · 2 Nov 2026

More than 1 in 3 AI agent skills had a security flaw. 76 carried confirmed malicious payloads. A skill is a folder with a SKILL.md file that teaches an agent like Claude Code a repeatable task. Great for productivity.

NeuralTrust on LinkedIn · 2 Nov 2026

In a benchmark of 1,348 tool poisoning test cases on 45 live MCP servers, the most exposed model hit a 72.8% attack success rate. No tested model refused more than 3% of the poisoned instructions.

NeuralTrust on LinkedIn · 2 Nov 2026

📰 This week in The AI Trust Letter Everything that happened this week for CISOs and Security Teams! • Anthropic Says Its Unreleased AI Turned Software Bugs Into Working Attacks on Chrome, Safari, Windows and Linux • Google and OpenAI Release New AI Models They Rate Among the…

Palo Alto Networks on LinkedIn · 31 Oct 2026

Frontier models reason. Decision models decide. Both need trustworthy inputs. As decision models power autonomous agents, securing the context informing their decisions becomes critical. Explore what's next for agentic AI ⬇️

F5 on LinkedIn · 31 Oct 2026

🚨 AI agents have the keys to your enterprise, but who’s watching the doors? ICYMI: As enterprises transition from simple AI prompts to autonomous AI agents, we enter the era of borrowed authority.

Singulr AI on LinkedIn · 31 Oct 2026

Awareness training still matters, but it only gets a company so far. This month we covered notetakers, MCP servers, skill files, pasted API keys, and more. If you only remember four things, make them these: • Connect an app for the job, then disconnect it when you're done.

Cranium AI on LinkedIn · 31 Oct 2026

The scariest AI in your company isn't the one on the roadmap. It's the model a team spun up in a sandbox last spring. The agent a vendor quietly turned on inside a tool you already pay for. The copilot nobody registered because nobody thought they had to.

Noma Security on LinkedIn · 30 Oct 2026

⚠️ A personal AI agent reportedly ran on a Pentagon network for two months. A record $6.3B left Binance wallets in about 20 minutes. Both started with something trusted inside the perimeter.

Noma Security on LinkedIn · 30 Oct 2026

🚨 A backdoored build of the Claude Code extension (v2.1.313) was distributed through Cursor’s marketplace, and updating won’t remove what it left on your machine. Here are 4 steps from Noma Senior Security Researcher Idan B.

Singulr AI on LinkedIn · 30 Oct 2026

A skill is a text file that teaches your AI how to do a specific job, like reviewing a contract or pulling together a weekly report. You download it, drop it in a folder, and your assistant follows it from then on.

Check Point on LinkedIn · 30 Oct 2026

What if your AI does something you never authorized? After an AI model unexpectedly attempted an external network connection, Check Point Research replayed the incident to investigate whether it was a backdoor or something else entirely.

Aembit on LinkedIn · 30 Oct 2026

🆕 ⚠️ WIRED published a pretty remarkable story this week about a defense contractor who reportedly connected a personal Claude Code account to a secure Pentagon environment and gave the agent credentials that let it write and deploy code with administrative access.

F5 on LinkedIn · 30 Oct 2026

If your worst cybersecurity habits came in a costume bag, they would look something like this. 🧛♂️ Swipe through for this year’s scariest enterprise IT nightmares—and no, we’re not talking about ghosts.

CrowdStrike on LinkedIn · 30 Oct 2026

AI adoption is accelerating. Stopping emerging AI threats takes elite expertise. As sophisticated threats hide in the noise of AI activity, distinguishing malicious behavior from legitimate use is increasingly difficult.

Check Point on LinkedIn · 30 Oct 2026

Using voice AI? Your transcript might not be telling you the whole story. 🎧 Attackers can manipulate audio so what the model hears and what your security sees aren’t necessarily the same. Audio Guardrails add protection where the interaction actually starts: the audio itself.

Quilr on LinkedIn · 30 Oct 2026

Every control your company has bought was built for humans. Agents are now the ones using them. Firewalls, security scanners, finance tools, marketing platforms: every one assumed a person was the actor.

Mindgard on LinkedIn · 30 Oct 2026

The bioweapons headlines were only part of the Kimi AI story. In the latest post in our Kimi series, we reveal how far the Apeiron jailbreak actually went.

Noma Security on LinkedIn · 29 Oct 2026

40,000+ Department of Homeland Security employee records leaked online. DDoSecrets states the data was scraped via DHS's Okta login. Okta reports that their own systems were not breached. The reality? Both are likely telling the truth.

CrowdStrike on LinkedIn · 29 Oct 2026

Prompt injection attacks are evolving. And adversaries are no longer just targeting what AI agents say, but how they think, plan, and act.

Zenity on LinkedIn · 29 Oct 2026

35 million euros or 7% of global annual turnover, whichever is higher. 💶 That’s the EU AI Act ceiling for prohibited AI practices. Now add GDPR, with fines up to 20 million euros or 4%. Add NIS2, with up to 10 million euros or 2% for essential entities.

Zenity on LinkedIn · 29 Oct 2026

Data residency used to be a mapping exercise. Agentic AI just made it dynamic. 🗺️ A single agent action can trigger a chain of tool calls, each one potentially routing through a different model inference endpoint, a different MCP server, in a different jurisdiction.

Zenity on LinkedIn · 29 Oct 2026

Identity tells you who an AI agent is. Access tells you what it can touch. 🔐 Neither tells you whether its last action was safe.

Vijil on LinkedIn · 29 Oct 2026

The idea of treating AI agent autonomy like a credit line is spreading. TechGig picked up the argument Vin Sharma makes that agents shouldn't get broad permissions on day one.

Pillar Security on LinkedIn · 29 Oct 2026

Our research team found a critical flaw (proposed CVSS 9.8) in OpenCode that let a malicious or hijacked MCP server run code on a developer’s Windows or WSL machine the moment they signed in.

AIR Security on LinkedIn · 29 Oct 2026

🚨 WARNING: If your AI agent set up a Solana project, someone else may already have your crypto wallet. AIR’s research lab found malware in the wild that targets AI coding agents. Here’s how it happens: 1. A developer asks their coding agent to build a Solana project. 2.

Pillar Security on LinkedIn · 29 Oct 2026

Who checks the chat template inside the open-weight model you just downloaded? Most GGUF models on Hugging Face ship with a chat template, a small Jinja2 program that runs on every inference call, and almost nobody inspects it.

Singulr AI on LinkedIn · 29 Oct 2026

The first time you open some AI browsers and agent apps, they offer to import your passwords, cookies, and history from Chrome. It's one click, and everything is already signed in when you start. Cookies are how a website remembers that you've logged in.

Giskard on LinkedIn · 29 Oct 2026

Within a few days this October, three critical security issues hit open source components that sit underneath a large share of production stacks, AI systems included: PHP2Shell, Node2Shell and a backdoored LiteLLM release.

Lasso Security on LinkedIn · 29 Oct 2026

Lasso security researcher Noy Pearl discovered CVE-2026-97693, a critical (CVSS 9.9) sandbox isolation failure in Yuxi-Know, a popular open-source multi-tenant agent platform built on LangChain and DeepAgent.

Harmonic Security on LinkedIn · 28 Oct 2026

Security frameworks help organizations structure their approach to risk. But what do they tell us about how employees should use AI? Our latest article explores six widely used frameworks: where they help, what it takes to satisfy them, and where teams need to go further.

Kong on LinkedIn · 28 Oct 2026

What if we told you that your API team already knows how to govern AI — they just haven’t been told yet? At API + AI Summit, Augusto Marietti , our CEO and co-founder, told Hard Coded that moving a token looks a lot like moving an API request. Same routing. Same rate limiting.

Palo Alto Networks on LinkedIn · 28 Oct 2026

Where Claude deploys, we defend 🛡️ Over 75,000 organizations depend on us to defend their digital enterprise. To deliver on that trust, we rely on Claude by Anthropic to accelerate our workflows, develop new capabilities, and uncover vulnerabilities faster.

Zenity on LinkedIn · 28 Oct 2026

Your company is probably already running AI agents. When one is manipulated or compromised, you’re not looking at a breached server. You’re looking at your company acting at machine speed, in ways no one approved.

Oasis Security on LinkedIn · 28 Oct 2026

“Is this workload who it says it is?” It’s harder to answer than it sounds. Humans have faces, fingerprints, and phones. Non-human identities have no biometric, so MFA doesn’t really apply, and what’s left is a secret.

Token Security on LinkedIn · 28 Oct 2026

National Institute of Standards and Technology (NIST) just published its 10 design tenets for agent identity. LinkedIn attention spans being what they are, here are several attempts to pique your interest.

Cranium AI on LinkedIn · 28 Oct 2026

Having written AI policies is no longer enough - you need verifiable evidence that your guardrails actually hold under pressure.

Rein Security on LinkedIn · 28 Oct 2026

Everyone in San Francisco is talking about sandboxes. 🧱 Our CEO Matan Bar-Efrat just spent two weeks there, and the topic came up everywhere, from the biggest tech companies to the big banks. His take: in practical terms, a sandbox is a way to manage compute for AI agents.

Lasso Security on LinkedIn · 28 Oct 2026

This month, an AI model submitted a false homicide tip to a police tip line during a test. A few weeks earlier, roughly 700 agent instances coordinated to breach Hugging Face while trying to solve a benchmark, and some of them edited logs to cover their tracks.

Singulr AI on LinkedIn · 27 Oct 2026

Nobody pastes a password into an AI tool on purpose. You copy an error log to figure out what broke, and there's a session token three lines down. You screenshot your terminal, and an API key is visible right above the line you actually wanted to share.

Aurascape on LinkedIn · 27 Oct 2026

Read the blog: https://hubs.ly/Q04Cw2xp0 Panasonic prohibited the AI tool. Data on approximately 334,000 people was uploaded anyway. For IT and security teams, the lesson is practical: written AI policies need controls that enforce them.

Noma Security on LinkedIn · 27 Oct 2026

What’s scarier than a haunted house? An ISO 42001 audit with no evidence. Policies? ✅ Risk methodology? ✅ Proof your agents actually followed them?... crickets... 🦗 ISO 42001 runs on a three-year cycle with surveillance audits in between.

Zenity on LinkedIn · 27 Oct 2026

Every security team has an AI acceptable use policy. Almost none of them can enforce it. 📄 Agents make thousands of decisions a day: tool calls, file reads, shell commands, outbound messages. A policy sitting in Confluence isn’t in that path.

HiddenLayer on LinkedIn · 27 Oct 2026

One in eight AI breaches is now linked to agentic systems, according to our 2026 AI Threat Landscape Report. Agents read files, call tools, and take actions, so your SOC sees events your queue has no rules for: a prompt injection that ends in an action, sensitive data leaving…

Pillar Security on LinkedIn · 27 Oct 2026

Pillar Research found a second code execution vulnerability in Unsloth, one of the most widely used AI fine-tuning libraries. Ariel Fogel’s first finding, in the article “Look, Don’t Load,” affected a beta version of Unsloth Studio.

Cranium AI on LinkedIn · 27 Oct 2026

🚨 NEW Blog: How to Stop Prompt Injection at Runtime Every few months someone announces they have solved prompt injection. Every few months a researcher gets past it in an afternoon. 😮 This is not a reflection on the engineering... it is a reflection on the problem.

Rein Security on LinkedIn · 27 Oct 2026

The PHP2SHELL incident feels like a breaking point. Whether frontier models create a vulnerability problem is no longer a question. It’s a fact. The real question is how you respond: more patching, or a real shift to prevention? We know where we stand.

NeuralTrust on LinkedIn · 27 Oct 2026

When agents do the work, where does the know-how live? Microsoft's 2026 Work Trend Index found active agents in Microsoft 365 grew 15x in a year. Every one of those agents is learning something: preferences, decisions, ways of working. The question is where that knowledge lands.

Thought leadership

Get in touch