A concise, fact-based update for security and risk professionals covering roughly the last 24 hours.

🔐 Core Security Intelligence

1) NSA, CISA, and international partners publish AI-in-OT security principles

What’s new
The U.S. National Security Agency (NSA), CISA, the Australian Cyber Security Centre, and other partners released a joint Cybersecurity Information Sheet titled “Principles for the Secure Integration of Artificial Intelligence in Operational Technology.” The guidance outlines how to safely introduce AI into OT environments that support critical infrastructure, focusing on governance, safety, and resilience.

Source: NSA, CISA, and Others Release Guidance on Integrating AI in OT

Why it matters
AI is starting to touch process control, industrial automation, and safety-critical systems. A misconfigured model, poisoned data, or AI-driven misdecision in OT isn’t just about data loss, it can impact physical processes, uptime, and human safety. This is one of the first major multi-agency documents that treats AI-in-OT as a distinct risk domain rather than a footnote to IT security.

Defenses / actions to consider

• Treat AI in OT as a separate architecture track: maintain a dedicated risk register and governance framework for AI components in industrial environments.
• Keep AI processing off the plant floor whenever possible: push sanitized OT data outbound to separate AI systems instead of embedding opaque models directly into safety loops.
• Design for fail-safe behavior: build guardrails and manual overrides so AI cannot unilaterally make or persist dangerous control decisions.

Expert insight
If you run critical infrastructure, this guidance is a strong hint of where regulators and auditors are heading. Expect future compliance regimes to ask specifically how you apply AI risk management inside OT, not just at the enterprise IT layer.

2) SandboxAQ launches AI Security Posture Management platform to tackle shadow AI

What’s new
SandboxAQ announced AQtive Guard AI-SPM, an AI security posture management platform aimed at discovering, assessing, and governing AI use across large enterprises. The tool discovers AI assets (models, agents, MCP servers), evaluates them for weaknesses such as prompt injection, data leakage, exposed credentials, and insecure dependencies, and provides continuous monitoring and policy enforcement.

Source: SandboxAQ launches new service to tackle shadow AI security risks

Why it matters
Most organizations already have “shadow AI” – untracked prompts, rogue agents, experimental models, or unvetted integrations quietly handling sensitive data. Traditional asset management and vulnerability scanners don’t see these AI-specific assets or their risk profile. An AI-SPM class of tooling is emerging to answer basic questions like: Where are we using AI? What data is it touching? What could be exploited?

Defenses / actions to consider

• Start building an AI asset inventory: models (internal and external), providers, agents, pipelines, and high-risk prompts, even if you don’t yet adopt a dedicated AI-SPM platform.
• Prioritize AI systems that touch secrets, production data, or privileged APIs for deeper assessment (prompt-injection exposure, data exfil paths, identity/credential handling).
• Align AI-SPM findings with existing risk registers, so AI weaknesses get triaged alongside “normal” vulnerabilities instead of falling into a separate, ignored bucket.

Expert insight
This is a sign that AI security posture management is on track to become as common as cloud or SaaS posture management. If you can’t answer “where is AI running and what can it break,” you’re not in control of your AI risk.

3) “Poetic jailbreaks”: research shows verse-based prompts can bypass AI guardrails at high rates

What’s new
New research covered by Malwarebytes described experiments where researchers turned malicious prompts into poetry to see if they could evade AI safety filters. Using poetic phrasing for high-risk topics (weapons, cyberattacks, privacy violations, misinformation), the team tested models from nine providers. For many models, poetic prompts significantly increased the attack success rate (ASR), with one open-source model showing a 62% ASR uplift. Even AI-generated poetry (not hand-crafted) achieved an average ASR around 43%, much higher than prose.

Source: Whispering poetry at AI can make it break its own rules

Why it matters
The findings expose a structural weakness in many safety stacks: they rely heavily on pattern matching and heuristic detection tuned for “normal” language. Once attackers wrap instructions in metaphor, rhythm, or unusual structure, those heuristics can fail badly. It also demonstrates model-on-model abuse, using one AI to automatically convert a library of known malicious prompts into high-evasion variants.

Defenses / actions to consider

• Do not rely on model-level guardrails alone, especially for high-risk use cases (code execution, infrastructure control, sensitive data access). Wrap outputs in downstream policy, filtering, and human review.
• Include “creative” adversarial prompts; poetry, coded language, oblique metaphors, in your internal red-teaming and safety evaluations.
• Consider using independent models or detectors trained specifically to spot obfuscated or stylized malicious content, not just blatant requests.

Expert insight
This research reinforces an uncomfortable truth: safety evaluations that only use straightforward, literal prompts are dangerously optimistic. Real attackers will happily get weird, artistic, or coded if that’s what it takes to slip past your filters.

4) Experian’s 2026 Data Breach Forecast names AI as the top cybersecurity threat driver

What’s new
Experian released its 2026 Data Breach Industry Forecast, calling out AI as “center stage” in the coming year’s cyber threat landscape. The report predicts more sophisticated attacks leveraging AI for identity fabrication, autonomous agents, shape-shifting malware, and deeper social engineering. Experian emphasizes that cybercriminals are often first to adopt new AI capabilities, using them to outpace traditional defenses.

Source: AI takes center stage as the major threat to cybersecurity in 2026

Why it matters
This is less about a specific CVE and more about strategic planning. When a major breach-response provider frames AI as the central driver of 2026 breach risk, boards and CISOs should expect questions from regulators, insurers, and auditors about their AI risk posture. It also hints at rising expectations that organizations will use AI defensively to keep up.

Defenses / actions to consider

• Update your enterprise threat model and risk register to explicitly cover AI-enabled attack patterns (AI-written phishing, automated recon, deepfake fraud, model abuse).
• Align breach simulation and tabletop exercises to scenarios where AI plays a starring role — both in the attack chain and in your response tooling.
• Work with legal, privacy, and comms teams to ensure your incident response plans cover AI-related breaches and AI-generated misinformation.

Expert insight
This kind of forecast tends to influence policy, regulatory guidance, and insurance underwriting. If AI is the headline risk for 2026, organizations that can show a credible AI security and governance program will be in a far better position than those treating AI as a side topic.

5) TÜV SÜD announces dedicated AI penetration testing services based on NIST, OWASP, and MITRE ATLAS

What’s new
TÜV SÜD announced new services focused on penetration testing of AI systems. Their offering evaluates LLMs and ML systems for threats such as prompt injection, model inversion, data poisoning, and adversarial examples. The testing is explicitly aligned with the NIST AI Risk Management Framework, OWASP LLM/ML security guidance, and MITRE ATLAS.

Source: TÜV SÜD tests IT-Security of Artificial Intelligence

Why it matters
Most organizations do not yet have mature internal capability to pentest models, data pipelines, or AI-enabled applications. External services that follow recognizable standards can help bridge that gap and provide evidence to auditors, regulators, and customers that AI risk is being actively tested — not just theoretically managed.

Defenses / actions to consider

• Add AI-specific testing to your standard security assessment lifecycle: before launch, after significant model updates, and after data-pipeline changes.
• Use frameworks like NIST AI RMF, OWASP GenAI/LLM, and MITRE ATLAS as the common language between your internal teams and external testers.
• Prioritize AI systems that impact high-risk processes (fraud decisions, safety-critical automation, sensitive data access) for the earliest and deepest testing.

Expert insight
We’re seeing the beginnings of a formal AI security testing ecosystem. Organizations that treat AI pentesting as normal, the same way they treat web app or network pentests, will be much better positioned when AI-specific regulations and due-diligence requirements tighten.

📊 At-a-Glance Summary

Categories: Cybersecurity News

Tags: AI Governance, AI Penetration Testing, AI Security, OT Security, Prompt Injection, Shadow AI