Those in cybersecurity who follow today’s most dangerous vulnerabilities say insiders, artificial intelligence (AI) prompt injecting, AI-powered phishing and ransomware as a service all top the list of concerns that warrant immediate attention from security leaders and practitioners. Interestingly, they note that AI itself has a few core weaknesses which leave it vulnerable, particularly to prompt-injection attacks.

Insider risk management  

Cybersecurity researcher, Dr. Jason Clark, presented his outlook in a recent Dark Reading webinar, sponsored by identity management firm, Okta. He says insider threats remain one of the most significant but often underestimated cybersecurity risks currently.

Insiders are those with legitimate access to systems. Both their malicious and unintentional actions can bypass traditional security controls typically in place to keep external attackers out, he says.

He adds that there are three distinct categories of insiders who may introduce risk. These include malicious insiders, unintentional insiders and business partners – vendors, contractors, service providers who may have privileged access. “It’s important to understand these distinct insider profiles,” he told those gathered for the webinar.

Types of attack then typically include fraud (this often requires collusion between two or more insiders), intellectual property (IP) theft and sabotage. IP theft can include programmers or salespeople taking company code and client lists to their next jobs. Sabotage is typically associated with a disgruntled employee. Using company resources for personal activities can open a company up to risks. And finally, espionage, the systematic, unauthorized collection and disclosure of confidential information, also earns a spot on Clark’s list of top threats.

“Each of these threat types really demands almost a tailored detection and prevention strategy.” He also adds that each demands a layered approach to the problem. 

Solutions: 

Safeguarding against threats from within requires proactive risk monitoring, clear policies and robust controls, Clark says. 

  • Implement a formal insider threat program. 
  • Establish structured processes to identify, assess and manage risks proactively. 
  • Adopt zero trust architecture wherever possible. Implement continuous identity verification, strict access controls and thorough user activity monitoring. 
  • Explore the many solutions available for user activity monitoring or UAM. 
  • Implement policies and training. Develop comprehensive guidelines, accountability standards for data handling, acceptable use and security practices.
  • Ensure that these policies are regularly updated and communicated. “Regular, engaging training is really essential,” Clark says.
     

“We want to basically try to verify each and every user and every access attempt, regardless of things like location or network,” he points out.

AI-enhanced phishing attacks  

One of today’s fastest-growing cyber threats, AI-enhanced phishing attacks go beyond traditional phishing’s manual tricks, allowing attackers to automate and scale their efforts “with really alarming precision,” Clark states.

He says the messages created are highly personalized, convincing and difficult to detect, even for well-trained users. Deepfake integration, including realistic impersonations of executives and colleagues, is particularly concerning. 

Clark points out that phishing today can also convincingly mimic legitimate internal communications by accurately referencing real projects, events, people and internal procedures. Automated campaign generation capabilities are also alarming, as are AI’s continuous learning capabilities.

“The combined effect of these innovations creates phishing attacks of unprecedented sophistication, scale and effectiveness – it really does challenge our traditional security awareness approaches,” he says.

Again, he adds, a comprehensive and multi-layered defense strategy is needed.

Solutions: 

  • Start with advanced security awareness training.
  • Establish clear verification procedures for sensitive requests. 
  • Deploy AI-powered advanced email filtering techniques and solutions. 
  • Deploy behavioural analytics and monitoring. Clark says this is similar to the behavioural analytics and monitoring used for insider threat detection. “We want to try and establish baselines of normal behaviour and then try to investigate deviations promptly,” he says.
  • Develop rapid incident response plans and capabilities. Develop AI-supported response workflows for quarantining compromised accounts and revoking access to minimize the damage caused by successful attacks.
     

“Time is really critical in containing these phishing-related breaches,” he says.

Prompt injection  

A relatively new and emerging attack vector is prompt injection. Clark says these attacks exploit vulnerabilities in AI systems by introducing maliciously structured inputs designed to deceive the AI model into performing unintended actions or disclosing sensitive information.

“AI systems inherently trust their inputs, making them really susceptible to malicious manipulation, especially in natural language processing (NLP),” he says. “Automated prompt injection techniques really amplify this threat by systematically probing AI vulnerabilities at scale.”

Prompt injection efforts can include prompting AI models with malicious inputs deliberately crafted to disrupt how AI systems make decisions. Context manipulation occurs when attackers try to feed misleading data and scenarios to the AI, causing it to misinterpret its environment. Direct instruction injection involves embedding commands that appear legitimate but are designed to bypass system safeguards.

The AI, he says, is going to trust its input by default. “Meaning they’ll process what they receive without enough scrutiny.” 

Interestingly, one question posed to the presenter asked if it is possible to fully secure large language models (LLM), “or are we just patching an inherently flawed architecture?” 

In response, Clark agrees that the flaw in question is a core weakness. “There’s really no true kind of isolation between prompts and system logic. It makes it more than a patchable issue, in my opinion,” he says. “We can do things like add filters and guardrails or monitoring, but at the end of the day I do think it’s an architectural flaw. True security I think, is going to require even more layered defenses – things like external rule engines come to mind. Identity verification comes to mind. Purpose-limited application programming interfaces (API) to enforce those boundaries.” He concludes by again saying it is currently a core weakness which requires more than just patching. 

Solutions: 

Layered, integrated strategies recommended in this case included three pillars of defense which Clark says reduce risk without compromising the benefits of AI adoption.

  • Examine input validation and sanitization, beginning with strong filtering that inspects and then cleans the inputs before the AI processes them.
  • Consider adding throttling and pattern recognition to catch probing attempts where attackers are trying to test the system’s limits.
  • Develop AI-specific incident response plans. “Who acts? Who gets shut down? What evidence is collected? And so forth,” he says. “Most importantly, in my opinion, we want to maintain the human oversight. AI, at this point in time, should support human judgement, not replace it, especially when there’s high-stakes vulnerabilities and decisions in play.” 
     

Ransomware as a service (RAAS) 

Finally, he says RAAS has fundamentally changed the ransomware threat landscape by lowering the technical barriers to entry for those launching advanced attacks.

Not only do the services operate under both subscription-based and profit-sharing models, but they also enable attackers with minimal technical expertise to “rent” ransomware tools and infrastructure through straightforward, user-friendly platforms. The platforms also provide continuous updates and technical support.

“This business model truly increases ransomware’s attack frequency, its scale,” he says. “What was once limited to maybe a handful of highly skilled attackers is now this scalable, subscription-based business model that has accessibility and profitability and is definitely constantly evolving, making it hard to mitigate.” 

He says RAAS is particularly dangerous because it operates like legitimate service providers, with flexible payment options and technical support no less, offering low-cost entry to even inexperienced cybercriminals who want to launch sophisticated ransomware campaigns. “They’re constantly refining their evasion techniques,” Clark says. “Really one of the most dangerous aspects of RAAS is how quickly it can scale.” 

He also warns of the service’s efforts to continuously improve, keeping RAAS ahead of its defenders. “These platforms are not static,” he says.

Solutions: 

Again, multiple, proactive strategies are recommended.

  • Start with threat intelligence. “The fight against ransomware really starts long before an attack even hits your network,” Clark states. He adds that companies need to stay informed about the latest tactics, techniques and procedures. To that end, he recommends subscribing to a trusted threat intelligence feed. “Learn from incidents that may be affecting your peers,” he says.
  • Implement continuous monitoring. “Regular threat hunting exercises are really critical.” He says many successful attacks are simply exploiting old, almost unknown vulnerabilities which haven’t been addressed.
  • Focus on robust endpoint protection as endpoints are often the first place that attacks strike. “Securing them is essential,” he says. “Timely patching is foundational.” 
  • Plan ahead. “Even with strong defenses it’s important that you plan for the worst,” he mentions.
  • Secure backup and recovery. Clark recommends the classic ‘3-2-1 strategy.’ That is for companies to keep three copies of their data on at least two different media types with at least one backup stored offsite.
  • Network segmentation. Enforce the principle of least privilege, ensuring that users only have access to what they need. Limiting access and isolating systems can significantly slow or even contain a ransomware outbreak, Clark says. “Segment your network, especially between departments or high-risk systems. Reduce the attacker’s ability to move laterally.”