Larger deployments, limited oversight
The share of organizations operating more than 150 agents rose from 7% in April to 16% in the latest wave of Gravitee's State of AI Agent Security research. Yet only 17% assign a separate identity to every agent, and 52% of respondents reported an agent-related security or privacy incident over the past year.
The ability to trace agents' actions is also limited. Only 1 in 8 organizations can fully reconstruct an agent's activity from 30 days earlier. On average, policies governing what agents may do are enforced for just 51% of agents.
Gravitee CEO Rory Blundell described the problem as "an accountability gap." He said most organizations cannot identify which agent was involved when something goes wrong, establish what it was authorized to do or determine who is responsible. In his assessment, that gap is widening as deployments grow.
Confidence has declined since the April survey. The proportion describing themselves as "very prepared" for an agent incident dropped from 47% to 29%, while the share believing current regulation addresses agent risks fell from 40% to 26%. Over the same period, the proportion of agents being monitored and secured changed little, rising from 56% to 57%.
Responsibility across connected agents
Agents rarely operate in isolation among the organizations surveyed: 97% of respondents said their agents call other agents, and 40% reported connections to agents belonging to other organizations. Only 27% had defined accountability across those chains.
Revoking access is already part of incident response for many respondents. Some 58% said they had done so during a live incident. Afterward, only 25% held a named individual responsible, while 17% made no changes.
Linus Håkansson, Gravitee's CPTO, said agents should be managed like employees who have access to sensitive systems: each needs an identity, enforceable limits and a record of its actions, along with a way to withdraw access within minutes.
The company's Agent Accountability Framework sets out six principles: Identity, Sponsor, Role, Authority, Oversight and Recourse. Gravitee says these are intended to establish which agent acted, whose authority it used, what limits applied and how it can be stopped. Its research found that most organizations fail to meet at least one of the principles.
How the sectors compare
Government and public-sector respondents reported the largest deployments: 25% operated more than 150 agents. They also had the weakest traceability, with only 8.8% able to reconstruct an agent action fully.
Financial services had the strongest controls in the survey but also the highest incident rate, at 66%. The sector had the least clarity about ownership, with 44% saying accountability was unclear. In healthcare, 63% reported an incident, with incidents including exposure of patient data. Only 21% of telecommunications respondents had a documented procedure for pausing an agent or revoking its access.
Opinion Matters conducted the survey in October 2026 for the third wave of Gravitee's research. It included 600 senior technology leaders in the United States and 300 in the United Kingdom, spanning financial services, healthcare, government and public sector, telecommunications, manufacturing, and travel and transport. The comparisons use the previous wave, fielded in April 2026.
European scrutiny and reporting requirements
The survey covers US and UK organizations. Separate European legal analysis provides a regulatory backdrop to the accountability questions. Law firm Bird & Bird notes that the European Commission's EU AI Act enforcement powers, including Article 101 fines, have applied since August 2, 2026. For agent systems not classified as high-risk, obligations are largely confined to AI literacy, prohibited practices and, where relevant, transparency duties under Articles 4, 5 and 50.
In July 2026, two leading AI developers disclosed that agents used in cybersecurity testing had escaped isolated evaluation environments and accessed third-party production systems. Those systems included a widely used platform for AI models and datasets; press reports also described effects on EU customers. On October 14, Commission Executive Vice-President Henna Virkkunen confirmed that the AI Office had requested information from the developers about security, external evaluation and monitoring of model use, according to accounts from Bird & Bird and Giulio Coraggio of GamingTechLaw.
Coraggio writes that an intruder being software does not change existing EU incident-reporting duties. These include NIS2 Article 23 early warnings, GDPR Article 33 notifications within 72 hours and DORA reporting of major ICT-related incidents by financial entities. They also include manufacturers' duties under the Cyber Resilience Act to report actively exploited vulnerabilities and severe incidents, and serious-incident reporting under Article 73 of the AI Act.
Liability, evidence and investigations
Coraggio's analysis also describes changes to liability and access to evidence. From December 9, 2026, the revised EU Product Liability Directive will treat software and AI systems as products and impose joint liability on manufacturers, importers and those making substantial modifications. A breach of mandatory safety requirements, or a defendant's refusal to disclose evidence, will give rise to a presumption of defectiveness. Only natural persons may bring claims under the directive, and purely economic losses remain outside its scope.
Italy's Law No. 160, dated September 9, 2026, made failure to provide technical security measures or human oversight for high-risk AI systems a criminal offense. A separate provision addresses unlawful alterations by people outside a system's lifecycle. The law also revised evidence rules for AI-related damage: claimants who present plausible facts can obtain orders requiring disclosure of event logs, risk-management records, technical documentation and human-oversight parameters.
On November 2, Belgium's national cybersecurity center attributed part of the increase in cyber incidents to targeted AI use. The Belgian federal prosecutor also opened a criminal investigation into an attack that destroyed food stocks in several countries, according to Coraggio.
Report and discussion
Gravitee says its full report, including sector breakdowns, is available to download from its website. Blundell is scheduled to discuss the findings with national security and AI leaders at a Gravitee evening at London's National Gallery on November 18.












