PwC finds attacks on AI systems top security leaders’ preparedness gaps

Cybersecurity budgets are set to grow, with AI a priority. But organizations’ reported defenses show a less confident picture. PwC’s 2027 Global Digital Trust Insights survey, published October 1, finds attacks on AI systems are security leaders’ biggest preparedness gap, while fewer than four in ten security, risk and operations leaders report fully formalized cyber continuity plans.

The survey covered 3,934 business and technology leaders across 71 countries. Help Net Security reported that fieldwork ran between May and July 2026. Its findings capture leaders’ assessments of readiness, spending and controls—not an independent test of how their organizations would withstand an attack. Different questions also drew on different leadership groups.

AI systems move ahead of familiar threats

Half of security leaders identified attacks targeting AI systems as one of their biggest preparedness gaps. That put AI ahead of cloud-related threats at 40%, third-party breaches at 34% and ransomware at 33%. The ranking concerns where leaders feel least ready; it is not a ranking of attacks they have experienced or measured losses.

PwC separately asked about AI-enabled attacks, a category that includes using AI to direct attacks as well as manipulating AI systems themselves. Security and risk leaders placed three threats at the top of their preparedness concerns:

• Autonomous botnets, 53%: networks of compromised devices that AI directs at scale.
• Adversarial attacks, 52%: subtle changes to what an AI system sees that cause it to answer incorrectly.
• Data poisoning, 52%: misleading records inserted into the data a model learns from.

PwC argues that improving AI capabilities make this gap more urgent. Its assessment is that frontier models can find previously unknown software vulnerabilities and exploit them through sophisticated attack paths with minimal human involvement. That is PwC’s assessment of the technology, distinct from the survey’s measurements of executives’ concerns.

More money, incomplete foundations

Among security and finance leaders, 84% expect cybersecurity budgets to increase, six percentage points more than last year. AI is among the leading destinations: 58% of security leaders rank it among their top cyber budget priorities. These are spending expectations, rather than evidence that the additional money has already produced stronger defenses.

Data protection and trust ranked among spending priorities for 51%, followed by securing against AI-enabled attacks at 46% and securing AI and autonomous agents at 45%. Yet companies reported implementing an average of only three of the seven data-risk measures PwC examined across their organizations.

49%Data classification policies fully implemented

About half have fully implemented policies for classifying data, the basic task of identifying sensitive information.

48%Data loss prevention fully implemented

Fewer than half report full implementation of data loss prevention across key channels through which data leaves the organization.

5%Every surveyed data-risk measure fully implemented

The share reporting full implementation of every measure fell from 7% last year.

Continuity planning presents another gap. Only 39% of security, risk and operations leaders reported fully formalized and integrated operational continuity plans specifically addressing cyber risks. Such plans document how to keep critical operations running or restore them after an attack. They address what happens during disruption, rather than only how to prevent an intrusion.

AI can help, but most leaders keep approval rights

For defensive AI, security leaders’ leading priorities were threat detection and alerting at 50%, fraud detection at 43%, and phishing detection and response at 42%. But interest in using AI does not translate into broad willingness to let agents—software that can take actions on its own—control the response.

Only 22% would authorize AI agents to carry out defensive actions without human approval. Another 38% would allow partial autonomy, while 36% prefer human-led execution with AI support across all actions. More than half named the technology’s reliability and maturity as a leading barrier to using agents for cyber defense.

Comfort with autonomous action also varies by task. Respondents were most comfortable authorizing agents for well-defined, repeatable work: enriching and connecting threat intelligence at 49%, quarantining or deleting phishing emails at 47%, and removing malware or repairing systems at 46%. The survey therefore identifies specific tasks where leaders are more open to autonomy, even while most stop short of granting agents unrestricted defensive authority.