Every Agent Is A New Identity

As AI agents are deployed across enterprise systems, they multiply identities and connections, making the management of their credentials, permissions and interactions as critical as managing those of human users — and, the World Economic Forum argues, likely more complex.

That assessment, in the Forum’s Global Cybersecurity Outlook 2026, published in January with Accenture, is the clearest statement in the report of how agentic deployment changes the security problem. Agents automate complex or repetitive activities with speed and consistency, but their integration challenges traditional security frameworks, redefines roles and processes, and raises fundamental questions about decision-making and the prioritisation of alerts.

The Global Cybersecurity Outlook 2026 is the fifth annual edition of the report, produced by the World Economic Forum in collaboration with Accenture and published in January. It draws on a survey conducted between 25 August and 1 October 2025, which returned 873 responses from 99 countries, normalised to 804 qualified participants from 92 countries, including 316 chief information security officers, 105 chief executives and 123 other C-suite respondents.

What can go wrong

Drawing on its earlier report AI Agents in Action: Foundations for Evaluation and Governance, the Forum sets out three failure modes that occur without strong governance.

Agents can accumulate excessive privileges. They can be manipulated through design flaws or prompt injections. And they can inadvertently propagate errors and vulnerabilities at scale.

The aggravating factor is tempo. The report notes that the speed and persistence of agents amplify these risks, and concludes that they require continuous verification, audit trails and robust accountability structures grounded in zero-trust principles that treat every interaction as untrusted by default.

The attack life cycle has already been automated

The report cites one disclosed case in support of its argument that the threat side is moving.

In November 2025, Anthropic disclosed a cyber espionage operation that demonstrated what the Forum describes as unprecedented use of AI across the entire attack life cycle — from reconnaissance and exploitation to data exfiltration. The report records it as the first confirmed case of agentic AI gaining access to high-value targets, including major technology companies and government agencies.

Its reading is measured. While generative AI is currently used primarily to enhance social engineering and reconnaissance, the emergence of autonomous AI agents capable of executing full-scale attacks signals a potential turning point.

Disclosure: Anthropic is the developer of the AI assistant used in preparing this article.

Governance has improved, and still trails deployment

The Global Cybersecurity Outlook 2026 records that the share of organisations with a process to assess the security of AI tools before deployment rose from 37 per cent in 2025 to 64 per cent in 2026.

The composition of that figure is where the exposure sits. Of all respondents, 40 per cent review AI tools periodically, 24 per cent assess once, 7 per cent do not know and 29 per cent have no process at all. The report notes that periodic review represents clear progress towards continuous assurance, but that roughly one-third still lack any process to validate AI security before deployment, leaving systemic exposures even as the race to adopt AI in cyber defences accelerates.

It adds that the market’s drive to adopt new AI features often outpaces security readiness, creating exploitable vulnerabilities.

Deployment is running well ahead of that governance. In the same survey, 77 per cent of organisations have implemented AI-enabled tools for cybersecurity: 52 per cent for phishing and email threat detection, 46 per cent for intrusion and anomaly response, 43 per cent for automating security operations, 40 per cent for user-behaviour analytics and 39 per cent for threat intelligence and risk prioritisation.

What the report says to do about it

The Forum sets out fundamental measures to secure AI at the infrastructure level. Data used in the training and customisation of AI models should be protected from breaches and unauthorised access. AI systems should be developed with security as a core principle, incorporating regular updates and patches. Robust authentication and encryption protocols should protect customer interactions and data.

On the operational side, the report identifies the balance between automation and human judgement as increasingly critical. AI excels at automating repetitive, high-volume tasks, but its limitations in contextual judgement and strategic decision-making remain clear, and over-reliance on ungoverned automation risks creating blind spots that adversaries may exploit.

Respondents to the Forum’s survey appear to share that caution. The requirement that AI-generated security responses be validated by a human before implementation is named as an implementation hurdle by 41 per cent, behind insufficient knowledge or skills at 54 per cent and ahead of uncertainty about risk at 39 per cent, insufficient funds at 36 per cent and an unclear business case at 33 per cent. The report reads the combination as evidence that trust remains a barrier to wider AI adoption.

The capability gap this creates

AI assurance is one of the sharpest dividing lines in the report’s comparison of resilient and non-resilient organisations.

Among highly resilient organisations, 83 per cent assess the security of AI tools before deploying them, with 71 per cent reviewing periodically. Among insufficiently resilient ones, 39 per cent assess at all, 20 per cent review periodically, and 55 per cent have no process.

Adoption of AI-enabled security tooling follows organisational scale: 93 per cent among large companies by revenue, 83 per cent among medium and 78 per cent among small, with the public sector at 70 per cent and non-governmental organisations at 45 per cent.

The report’s conclusion on that distribution is that differences in capability significantly increase systemic exposure within interconnected supply chains, because adversaries can target less-protected partners to reach high-value organisations downstream. More than half of all respondents, 54 per cent, identified limited knowledge and skills as the key obstacle to adopting AI-driven solutions for cybersecurity.