By Cha’Von K. Clarke-Joell AI Psychology, Root Ethics Strategist, human-centred AI governance practitioner, author and educator. Founder and CEO, CKC Cares Group
For years, cybersecurity has repeated a familiar phrase: people are the weakest link. I have never been particularly comfortable with it. The phrase isolates the person from the conditions under which that person is being asked to make a decision. Those conditions are increasingly part of the problem.
A convincing email no longer needs awkward grammar. A voice on the telephone can sound familiar. A request can contain accurate contextual information; an account can be authenticated and still misused; and an AI agent can have legitimate permissions but still act in ways that create risk. The challenge has moved from whether someone can identify something suspicious to whether we are giving people a reasonable chance of knowing what to trust.
AI-enabled social engineering is not just a cybersecurity problem requiring a cybersecurity solution. It is also a problem of identity, psychology, trust, AI capability, organisational culture, governance, access
control and human judgement.
Cybersecurity operates inside a wider digital polycrisis, where AI, privacy, identity, organisational behaviour and security risks interact instead of remaining separate. Treating each as an isolated discipline makes it harder to see how one weakness amplifies another.
A synthetic voice can become an identity-verification problem when voice or familiarity is used, formally or informally, as evidence of identity.
Finally, think about how an autonomous agent with legitimate credentials can become an access control and governance problem. Authentication answers only one question: does this entity have valid credentials? It does not answer whether the agent should be taking this action, in this context, for this purpose, at this time. That raises questions about permission scope, least privilege, monitoring, escalation, accountability and who has authority to stop the agent.
Mandiant’s M-Trends 2026 found that exploits remained the most common initial infection vector in its 2025 investigations, while highly interactive voice phishing rose to second place. Google Threat Intelligence Group also reported adversaries moving from basic prompting towards agentic AI workflows and AI-enabled automation. In an observed incident in the second quarter of 2026, threat actors compromised a cloud resource, then planned, built and executed an agent-enabled mass
credential harvesting campaign in under six hours.
Mandiant is also explicit that the vast majority of successful intrusions it investigated still stemmed from fundamental human and systemic failures. ENISA’s 2026 Threat Landscape continues to identify established threats, including ransomware and vulnerability exploitation, while noting the growing use of AI to support malicious operations. The point here is that familiar weaknesses are now operating in a changing environment, and that changes what we mean by the human attack surface.
If a synthetic voice sounds like a senior colleague, the request contains credible information, the caller understands internal processes and the interaction creates urgency, telling an employee to be more vigilant is not an adequate security strategy. The vulnerability may not be the human being. Instead, the conditions under which people must make the judgement may actually be the hidden vulnerability.
Today, we face a related challenge: establishing the authenticity of the interaction itself. Is the person actually who they claim to be? Is the instruction genuine? Is an authenticated account behaving as its legitimate owner intended? Is the autonomous system acting according to the purpose and authority it was granted?
I think of this as synthetic trust: a condition in which the cues we normally use to decide whether a person, message, instruction or interaction is trustworthy can themselves be artificially produced,
reproduced or manipulated. Familiarity, authority, fluency and contextual accuracy are useful trust cues, but no longer enough on their own.
This has implications for threat intelligence, because a signal is not the same thing as understanding. Security may detect unusual authentication behaviour. HR may know that a person has recently changed role, a service desk may have received an unusual reset request, and a supplier may hold another piece of the picture. Each observation can be accurate, yet the organisation as a whole may not realise or understand what is happening.
I describe this as Distributed Sensemaking Loss: knowledge exists across the system, but is not connected quickly enough to create shared understanding and support sound judgement. As attack cycles become faster, the time available to assemble that understanding becomes smaller.
The next challenge for threat intelligence may actually be less about collecting more signals and more about improving an organisation’s capacity to make sense of them. AI can increase the speed, scale and credibility of some forms of attack, which makes the distance between detection and judgement matter. And so, the distance between teams that hold part of the picture will matter more.
Threat intelligence will need to ask:
• Who can interpret the signals?
• Who can connect technical indicators with organisational context?
• Who has authority to challenge, escalate or stop an action?
• What happens when an incident does not fit the playbook?
• How easily can an employee report uncertainty without fearing that they will be blamed for raising a false alarm?
Alongside established indicators of compromise, organisations may also need to become more alert to indicators of manipulated trust. These might include an unexpected appeal to authority, an unusual request to bypass verification, a change in normal communication behaviour, an authenticated identity acting outside its expected context, or an AI agent taking actions that are technically permitted but inconsistent with the purpose for which access was granted.
This is an argument for recognising that technical controls operate inside human and organisational systems.
Security awareness should reflect that reality. Providing opportunities to practise ambiguous situations beyond the obvious phishing examples helps to build confidence and proactive responsiveness.
Reporting uncertainty should be normal, and escalation routes should be clear before an incident occurs. As synthetic media, AI-assisted attacks and autonomous systems become more common, trust becomes part of the attack surface. The response should not be to distrust everything, or to place an even greater burden on individual employees to detect the undetectable.
A more useful question is whether we are building organisations capable of making good decisions when familiar signals of trust become less reliable. Threat intelligence can help us see what is happening. Human judgement, supported by good governance and organisational sensemaking, determines what we do with that knowledge while there is still time to act.
Contributor Bio
Cha’Von K. Clarke-Joell is an AI Psychology and Root Ethics Strategist, human-centred AI governance practitioner, author and educator. She is the Founder and CEO of CKC Cares Group, where her work examines the intersection of artificial intelligence, human judgement, organisational behaviour, governance and digital risk. Her work focuses particularly on the human and systemic conditions that shape how organisations adopt, govern and respond to emerging technologies. She is the author of The Digital Polycrisis, was recognised as one of the 100 Brilliant Women in AI Ethics in 2023 and a Top Ten AI Speaker for 2026. Cha’Von also teaches in higher education and has previously worked in privacy regulation, including in innovation, AI and emerging technologies.
Latest News
