Only a quarter of enterprise threat intelligence teams have deployed agentic artificial intelligence in live cybersecurity operations, as concerns about reliability, oversight and unintended consequences constrain adoption, new research from Flashpoint shows.The Enterprise Threat Intelligence Report 2026, released on Tuesday, found that 25% of organisations surveyed were using AI agents in production, while another 20% were conducting pilots. A further 35% planned adoption within 12 to 24 months, and 20% remained cautious or had no immediate deployment plans.
The findings highlight a divide between growing investment in AI-assisted defence and security professionals' willingness to give software systems authority to act independently. Unlike conventional tools that produce recommendations, agentic systems can undertake sequences of tasks and, when authorised, make changes to security controls without requiring approval at every step.
Respondents identified inaccurate decisions, leakage of sensitive information, insufficient audit trails, regulatory obligations and uncertain costs among the obstacles to wider use. Their reservations were particularly pronounced where agents could alter network settings, block legitimate activity or take other actions affecting business operations.
Flashpoint said organisations were evaluating the technology against demanding standards for accuracy, measurable value and human supervision rather than treating autonomy as an objective in itself. Some were restricting agents to gathering information or recommending responses, leaving consequential decisions to security analysts.
Josh Lefkowitz, Flashpoint's co-founder and chief executive, said security leaders recognised the opportunity presented by autonomous tools but wanted evidence of improvements to their work, trustworthy outputs and clarity about where human judgement remained necessary.
The research identified several applications already being tested or deployed, including retrieving threat intelligence, prioritising alerts, analysing security monitoring data and identifying suspicious network traffic. More active uses included blocking internet protocol addresses and updating lists of prohibited connections.
Pilot programmes also covered detection and validation, generating investigative context, querying threat intelligence platforms and mapping activity against the MITRE ATT&CK framework, which catalogues adversary techniques. Reversible response actions attracted interest because teams could assess their effects and undo changes when necessary.
One retail-sector respondent would consider only read-only AI assistants, citing the risks of granting software permission to modify systems. Another participant warned that a confidently incorrect agent could be more damaging than a slower system, illustrating why speed alone was not sufficient justification for deployment.
The caution comes as security departments confront mounting demands to detect AI-enabled attacks. Flashpoint found that 65% of respondents regarded excessive alerts and false positives as a leading operational challenge, while 55% cited inadequate analyst capacity and 50% identified difficulties integrating different security products.
AI-related threats ranked highest among areas where organisations wanted to expand intelligence coverage. Priorities included detecting deepfakes and synthetic media, monitoring faster exploitation of software vulnerabilities and identifying emerging attack techniques supported by artificial intelligence.
Budget expectations nevertheless indicated continued investment. Half the respondents anticipated higher threat intelligence spending over the next year, including 35% expecting increases of at least 10%. Another 40% expected budgets to remain broadly unchanged, while none forecast reductions.
Additional spending was being directed towards AI-related threat coverage, intelligence feeds and improvements to existing programmes. The figures suggest that reluctance to grant agents operational independence does not necessarily mean organisations are cutting expenditure on technology or intelligence capabilities.
Separate research published by Check Point this year found AI agents operating in pilot or production environments at 64% of surveyed organisations, although that study covered enterprise AI use more broadly rather than the specific threat intelligence workflows examined by Flashpoint. The differing scopes make direct comparisons inappropriate.
Flashpoint also tracked more than 22 million posts by threat actors discussing, sharing or advertising AI toolkits during the first half of 2026. The activity included material associated with phishing, malicious software development and other cybercrime techniques, reinforcing the pressure on defenders to improve their understanding of AI-enabled threats.
For organisations considering deployment, the report identified reviewable outputs, clearly defined permissions and documented approvals as practical safeguards. It also emphasised the importance of recording agents' actions and limiting access to systems and information according to operational requirements.
The study was commissioned by Flashpoint as organisations prepared their threat intelligence priorities and budgets for 2027. Its findings distinguish between tools already operating in production, controlled experiments and intentions to adopt technology over a longer planning period, providing separate measures of actual implementation, limited testing and planned investment across the organisations surveyed.
Topics
Technology