Mock review exercise 3
- Type
- AI R&D pilot in workplace cybersecurity
- Duration
- 18 months
- Applicant
- PhiShield with partner company NorthBridge
- Purpose
- email security and staff cyber-risk management
- Output
- phishing susceptibility score plus behavioural labels
- Review status
- submitted for ethics review after a pilot incident
Project summary
PhiShield is an AI-supported cybersecurity system designed to reduce phishing and business-email compromise in large organisations. It combines email metadata, click behaviour in simulated phishing exercises, training-test performance, and workplace communication patterns to generate a phishing susceptibility score for employees.
It also produces behavioural labels such as “authority responsive”, “urgency reactive”, or “procedurally compliant”. The applicant states that the system is intended as a decision-support tool for security awareness and risk mitigation, not as an autonomous decision-maker.
Technical design
The model is trained on historical enterprise email events from three multinational firms. It uses features such as response patterns, deviation from workflow, phishing-test behaviour, escalation habits, and linguistic markers in replies.
The system produces a risk score from 0 to 100 and a short natural-language explanation based on contributing features. Confidence scores are logged internally but are not shown to affected employees. PhiShield states that protected attributes are not used directly, but job role, seniority, language, and reporting structure may influence the model.
Pilot incident
At NorthBridge, a new department manager excluded a senior executive assistant with twenty years of service from the Executive Committee after the system flagged her for authority bias. The manager explained that people who strongly follow leadership instructions were considered more likely to comply with a spoofed executive request.
The employee had not been informed in advance that behavioural security scores could affect committee membership, received only a brief managerial explanation, and was given no appeal route. They interpreted the incident as a stigmatizing judgment that mispresented loyalty and procedural discipline as a defect.
When reporting pilot results back to PhiShield, NorthBridge security department included the following statement:
“Business email compromise and targeted phishing cost our industry billions annually, and the tactics are evolving faster than our brains can keep up with. Information security is an entirely new risk that cannot be adequately understood by senior staff trained decades ago in a different environment. This software is working exactly as intended. In just the pilot phase, we’ve seen a 19% reduction in successful phishing-test clicks.”
Dashboard example
| Indicator | Behaviour | Explanation |
|---|---|---|
| Primary indicator | Authority Responsive | User exhibits high compliance with leadership directives. |
| Other indicators | Rapid Response Time | Response time to executive emails is in the 95th percentile (Avg: < 3 mins). |
| Other indicators | Workflow Adherence | Extremely low frequency of deviation from formal workflow. |
| Other indicators | Linguistic Markers | High use of deferential linguistic markers in replies (e.g., “Right away,” “As you requested”). |
| Escalation patterns | Low | Rarely transfers unusual requests to IT or other colleagues. |
Reviewer’s task
- Using the Checklist, identify serious and complex ethics issues and map them under the relevant ALTAI requirements.
- Using the Portfolio, propose up to 8 measures to address these issues.
- Detect potential ethics tension(s) in between the ALTAI ethics requirements.
Source: A. Prégent and A. Grinbaum, Developing Competency in AI Ethics: A training framework. arxiv:2610.13319
