The Hacker News | Expert Insights

AI agents have become incredibly useful across many industries by their ability to reach inboxes, documents, and financial information, then search, summarize, or act on what they find. But that access creates a data-in-use problem: conventional encryption protects information in storage and transit, but an agent generally needs it decrypted in memory while processing it, where sensitive material can be exposed to application code, logs and debugging systems, infrastructure operators, or a compromised host. Conan Yu's work offers one practical response to that problem. He is co-founder of Rena Labs , which develops infrastructure for confidential AI training and inference using trusted execution environments (TEEs), hardware-isolated environments designed to limit the surrounding host's access while code and data are processed. Over roughly two years, Yu has worked across hardware and cloud TEE deployments, privacy-preserving financial-data analysis, and private AI inference. He a...

More than half of IT & cybersecurity professionals who experienced a breach in the past 12 months say they were told to keep it confidential, even when it was reportable. That finding comes from the 2026 Bitdefender Cybersecurity Assessment , a study Bitdefender has run for several years running. Disclosure rules have expanded significantly since the question was first asked in 2023, but the pressure to hide breaches continues. How Many Organizations Hide Breaches? Roughly half of 1,200 IT and security professionals surveyed reported a breach or security incident in the last 12 months. Of that group, 55.2% said they had been asked to keep a breach confidential even when it should have been disclosed. That's a snapshot. Here's the trendline: In 2023, 42.0% of respondents said they'd been asked to keep a breach quiet. By 2025 that number had climbed to 57.6%. In 2026 it settled back to 55.2%, which appears to be a plateau, not a reversal. IT & security pr...

Check your expanding identity attack surface Identity has become the new perimeter, and attackers know it. According to new research from CrowdStrike, 80% of modern cyberattacks are identity-driven , leveraging compromised credentials.  This substantial volume is driven by the widespread use of well-known directories such as Active Directory, and the vulnerabilities associated with compromised privileged accounts that are overprivileged or unmanaged, not network intrusion.  The explosion of privileged credentials and lack of adequate visibility or oversight can be attributed to growth, whether due to migration to the cloud, mergers and acquisition, the increased use of contractors, or simple organic business growth. This growth outpaces the ability to adequately govern the environment, leaving exploitable vulnerabilities.

Most security teams know that dwell time matters. The harder question is what to do about it. Dwell time is the period between an attacker gaining access and the security team containing the threat. During that window, a threat actor has time to learn the environment, steal credentials, move between systems, and reach sensitive data. For years, security teams have tried to reduce this window by adding more detection tools. The problem is that more alerts do not necessarily mean faster detection.  A recent industry incident response report puts the global median dwell time at 14 days, up from 11 the year before, quietly reversing a run of steady improvement that had held for close to a decade. The better way to think about it is as an operational problem. Two numbers matter most. Mean time to detect (MTTD) tells you how quickly the team recognizes a real threat, while mean time to respond (MTTR) tells you how quickly the team investigates and contains it. An AI native SIEM...

As organizations deploy AI coding agents across large monorepos and microservices environments, a fundamental problem emerges: the model may be capable of making the change, yet still lack the organizational context required to make the right change safely. A developer can ask an AI coding agent to deprecate an API field, update an authentication flow, or modify a service interface. The agent can inspect the code available on the developer's machine and search for references. What it may not know is that the field is consumed by four other services across separate repositories, that one of those services belongs to another team, or that the same field eventually carries sensitive data into a third party integration. This is not simply a context window problem. It is a code context problem: providing AI agents with accurate, current, organization wide evidence about how software actually behaves. One emerging approach is to generate that evidence directly from source code us...

AI coding agents are already inside engineering organizations. The problem security teams need to solve is not only that AI-generated code might be vulnerable. You already have ways to catch that: code review, CI, SAST, dependency scanning, and production monitoring. The real problem is that tools such as Claude Code, OpenAI Codex, Claude Cowork, and GitHub Copilot are becoming extensible agent runtimes. Skills, plugins, hooks, repository instructions, and MCP servers can influence what the agent reads, which tools it selects, what commands it runs, and where enterprise data is sent. Most AI governance programs stop at approving the application. Very few can tell you everything that has been installed inside it. That is the supply-chain gap. What changed: Third-party components are no longer participating only at build or deploy. They are participating in the agent's decision loop. From coding assistant to agent runtime The first generation of coding assistants mainly...

In eleven days, on September 11, manufacturers of products with digital elements sold into the European Union have to tell a regulator within 24 hours of learning that a vulnerability in one of their products is being actively exploited, with a fuller account due at 72 hours. I have a decent idea what the next eleven days look like inside most of those companies, having spent close to thirty years watching software organizations get ready for a date on a calendar. There will be a spreadsheet of products and owners that somebody builds over a weekend, a notification template that goes to legal for review, probably a consultant on a two-week engagement. It will mostly work. By September 10, the majority of them will be able to file inside 24 hours, and they will be right to feel relieved about it, because filing on time is exactly what the regulation asks, and it is not a trivial thing to arrange. What I would gently point out is that almost none of them will come out of the exercise ...

Cybersecurity teams have no shortage of threat data: New vulnerabilities are disclosed, malware is discovered, attack campaigns are analyzed, and manufacturers, CERTs, and security agencies continuously publish indicators of compromise (IoCs), security advisories, and other technical information. For operators of critical infrastructure, however, collecting this information is not even the most challenging part. Security teams still need to determine whether a threat is relevant to their environment, which assets may be affected, and what the observed activity actually means in the context of an operational network. In the energy sector, that requires knowledge extending beyond enterprise security and into the protocols, equipment, and processes that keep power systems operating. A suspicious packet in an office network is one thing. Understanding whether communication between an engineering workstation and a protection device using IEC 61850 represents expected maintenance activi...

Many AI systems depend on input signals that teams cannot fully inspect or explain. These opaque sources reduce visibility into the data paths that influence model behavior. Engineers lose provenance records, limiting the diagnosis of abnormal outputs. This complicates the work of security teams that need clear records of what influenced a model at any point in time. Verifiable search data offers a stable alternative. It gives teams an input they can examine, store, and reproduce in controlled conditions. Engineers can compare model behavior against information that was publicly accessible at the time a result was produced, rather than depend on hidden internal signals. This article outlines why verifiable search data gives AI and security teams the clarity required to maintain operational control. Why Traceability Matters in AI Systems Traceability lets teams follow an input from its origin through each processing step. When every stage can be inspected, engineers can review...

Artificial intelligence is everywhere now. From automated code completion to autonomous infrastructure management, AI tools and AI agents help DevOps speed up deployment cycles and change how development teams operate in general. At the same time, this rapid adoption of AI has created a reality that is hard for security teams to ignore: as with the growth of AI capability within the software development life cycle, the attack surface also grows. In 2025, there were 68 AI-related incidents recorded across major DevOps platforms according to the 2026 DevOps Threats Unwrapped Report . In the first half of 2026, the number of AI-related incidents visibly grew — research from GitProtect Lab tracked 84 AI-related incidents in six months alone. Thus, comparing the first half of 2026 to the same period in 2025 shows that AI-related incidents in development environments have nearly tripled. What do DevOps and DevSecOps say about AI incidents in general? According to GitProtect Lab 's surve...

Penetration testing is intended to help organizations identify weaknesses before attackers can exploit them. Once testing ends, findings must be documented, reviewed, formatted, delivered, assigned, tracked, remediated, and eventually retested. In many organizations, each of those steps happens in a different system and depends on a manual handoff. Testers work in one set of tools. Reports are assembled in Word or spreadsheets. Findings are delivered through PDFs. Security teams recreate them in ticketing systems. Engineering teams update remediation status somewhere else. Retesting is coordinated through email or meetings. By the time the right owner receives the information needed to act, days or weeks may have passed. At PlexTrac , we see this as one of the largest operational gaps in modern offensive security: organizations have invested in finding vulnerabilities, but the process surrounding the pentest has not kept pace. The next phase of pentest modernization is removin...

Identity is the attack surface now. Most identity governance and administration (IGA) programs still run on manual certifications, static role models, and quarterly reviews that go stale the day someone signs off on them. That's not a compliance inconvenience for a CISO. It's a structural gap. Attackers don't wait for the next recertification cycle, so identity risk detection can't either. Autonomous identity governance turns IGA from a periodic, human-driven exercise into something that runs continuously, watching real usage, learning what normal looks like, and acting on deviations before they turn into incidents. That autonomy applies across every identity and entitlement placed under governance, continuously reassessing access as usage, roles, and risk signals change. Three things are colliding to force this shift. Identity sprawl across cloud and SaaS environments has grown past what manual reviews can realistically handle, service accounts and non-human identi...

A security leader at a global finance company told us recently that his team discovered three times more AI tools running in their environment than IT had approved. Nobody had smuggled them in. Employees had simply pointed agents at their work, and the agents brought their own tools with them. That conversation is not unusual. It is the conversation. Over the past year, in customer discussions across finance, healthcare, manufacturing, and government, the same four struggles come up so consistently that we have started treating them as the shape of the problem itself. Every company effectively hired a second workforce this year, human workers and agentic workers side by side, and the agentic workers never went through onboarding. No handbook, no scoped credentials, no acceptable-use policy they can actually read. Here is what teams are struggling with, what our research says about why, and what closing each gap actually requires.

When it comes to email security software, the cybersecurity industry has become very good at one thing: scoring the message that's already in the inbox. We've gone from blocklists to signature matching to behavioral ML, and each generation of email security innovation was a genuine improvement over the last.  But the losses keep climbing anyway. The median time it takes someone to click a phishing link is just 21 seconds after opening the email, and another 28 seconds to hand over credentials or payment data. That's under a minute start to finish. That number hasn't moved much in years despite everything we've thrown at the problem. It's time for a different approach. But first, we need to understand modern challenges. We're filtering messages, but attackers are running campaigns

Somewhere right now, malware running on a compromised machine is checking in with an AI model, asking it for a new version of itself. Google's Threat Intelligence Group caught this happening in late 2025. The malware, nicknamed PROMPTFLUX, does this every hour it runs, and each version comes back looking different from the last. By the time a security tool learns to recognize it, it has already changed shape again. This isn't a rare glitch or a lab experiment. It's a preview of how a growing share of attacks work today and why legacy SIEM platforms built to detect known patterns are starting to fall behind. Legacy security tools were built to detect, not to adapt A SIEM is a system that collects logs from every part of a company's network and looks for signs of an attack. For years, it worked like a security guard with a very long memory. The guard learns what a break-in looks like - a certain kind of file, a pattern of behavior, a code signature, and watches fo...

We had an enterprise customer tell us their entire AI footprint was Copilot. That was the whole answer. One tool, one line item, done. We ran the first scan. Copilot wasn't even close to number one. Claude was. OpenAI came in second. Copilot was third. Nobody on the security team knew because Claude doesn't show up the way a SaaS app used to show up. There's no single login screen, no single admin console, no one place to look. That's the part most security teams miss. Claude isn't one surface. It's six. The six places Claude actually runs Claude Enterprise and Connected Apps. This is the surface everyone pictures: employees typing into Claude, OAuth'd into Google Drive, GitHub, Slack, and Jira, asking Claude to act on what's inside. The audit log shows that a connection happened. It does not show what got pulled into the prompt or what came back out. A finance analyst can drop a quarter of board materials into a conversation in ten second...