The Prudential Authority has told South African financial institutions that the window between a vulnerability becoming known and being exploited is closing, and that it will now supervise how they act on that, not whether they are aware of it. Its communication of 24 April 2026, Preparing for Frontier AI and AI-Accelerated Cyber Risk, applies to every supervised financial institution. It states that frontier AI “materially compresses the time between vulnerability discovery and exploitation” and that the PA’s supervisory focus is shifting “from awareness of AI-accelerated cyber risk to execution, operational resilience, and effective decision-making under compressed timeframes”.
The communication is calm in tone. It says plainly that these developments “do not signal an immediate crisis for the South African financial sector”. But its practical expectations are demanding, and they apply well beyond banking.
What does the Prudential Authority actually expect?
Frontier AI refers to the most capable general-purpose AI models available at a given time. The PA notes that recent public coverage of one such model preview showed advanced models autonomously discovering high-impact software vulnerabilities and generating working exploits. Its conclusion is that attackers will increasingly find and weaponise weaknesses “at machine speed”.
The supervisory expectations that follow are specific:
- Prioritise by exploitability, not volume. A backlog of thousands of open findings ranked by generic severity is not a risk picture. The question is which weaknesses can actually be reached and used.
- Validate exposure continuously. The PA names applications, dependencies, third-party connections, systems and automated service accounts.
- Detect and respond at machine speed, through appropriate automation supported by strong, ongoing identity and access controls.
- Treat N-day vulnerabilities as urgent, particularly in legacy and third-party environments.
- Use AI defensively and securely for investigation, triage, containment and remediation.
- Make internal AI use visible. AI usage across development and operational environments should be governed and monitored to detect unsafe behaviour early.
An N-day vulnerability is a weakness that has been publicly disclosed and for which a fix or advisory exists, but which has not yet been applied. Most real-world breaches use them. The March parliamentary reply on Land Bank, which we discussed in our analysis of internet-facing systems, is a reminder of what one unpatched, exposed server can lead to.
Why is an annual penetration test no longer sufficient on its own?
A penetration test is an authorised, simulated attack against defined systems to find exploitable weaknesses before an attacker does. It remains essential. What has changed is the interval between tests.
Many regulated firms in South Africa still test their external perimeter once a year, often timed to an audit cycle. If exploitation can now follow disclosure within days, an annual test describes the environment as it was on one date. For the remaining months, the organisation relies on patching discipline and monitoring it has not independently validated.
The PA’s language points to a different model:
- Continuous external monitoring of the attack surface, so new exposures are visible when they appear rather than at the next test.
- Periodic deep testing, at least annually and after significant change, by testers independent of the team that built the systems.
- Targeted retesting whenever a critical vulnerability is published that affects a technology the organisation uses.
- Identity testing, including service accounts, because automated accounts with broad privileges are now an explicit supervisory concern.
This is not a call to buy a new platform. It is a call to stop treating testing as an annual event and start treating it as an ongoing control.
What does the PA expect of boards?
The communication is unusually direct on governance. AI-accelerated cyber risk is described as a board-level and executive responsibility. Boards and senior management are expected to ensure that:
- decision rights for containment and recovery are pre-defined, so nobody is waiting for a meeting to approve isolating a system;
- risk appetite is reviewed against compressed attack timelines, so a 30-day patch window is a decision rather than a default; and
- management can detect and respond at the required speed and scale, with evidence.
The first point matters most in practice. In many incidents the slowest step is not technical. It is the time it takes to find someone with the authority to take a revenue-generating system offline.
How does this relate to Joint Standard 2 of 2024?
The PA is explicit that the communication does not replace or reduce existing obligations. It describes the Joint Standard on Cybersecurity and Cyber Resilience, in force since 1 June 2025, as “the cornerstone of regulatory expectations for governance, detection, response, recovery, and timely reporting of incidents”. Institutions significantly aligned with the Joint Standard’s intent, it says, will be better positioned to withstand what is coming.
For institutions inside its scope, that means the frontier AI communication is a lens on existing requirements. For firms outside it, such as law firms, medical practices and professional services firms, the Joint Standard is still a useful benchmark. The attackers using these tools do not check which regulator supervises their target.
What to do this quarter
- Rebuild the vulnerability backlog around exploitability. Separate internet-reachable and actively exploited weaknesses from the rest, and set short deadlines for that subset.
- Introduce continuous external monitoring of your attack surface between formal tests.
- Schedule an independent penetration test if your last external test is more than twelve months old, and agree a retest trigger for critical disclosures.
- Inventory service accounts and remove standing privileges that no longer have an owner or purpose.
- Write down containment decision rights. Name who can isolate which systems, without further approval, out of hours.
- Review your patch windows with the board and record them as risk-appetite decisions.
- Map AI use in development and operations, including coding assistants and automation tools.
Frequently asked questions
What did the Prudential Authority publish on frontier AI?
On 24 April 2026 it issued Preparing for Frontier AI and AI-Accelerated Cyber Risk to all supervised financial institutions. It says frontier AI materially compresses the time between vulnerability discovery and exploitation, and that supervision will focus on execution and resilience rather than awareness.
Does the communication create new regulatory obligations?
No. The PA states that existing obligations are neither replaced nor reduced, and that the Joint Standard on Cybersecurity and Cyber Resilience remains the cornerstone. What changes is the evidence supervisors will expect to see.
What is an N-day vulnerability?
It is a publicly disclosed weakness for which a fix or advisory exists but has not been applied. The PA expects institutions to treat these as urgent, especially in legacy and third-party environments.
We are not a bank or insurer. Does this apply to us?
Not formally. But the threat it describes is not sector-specific, so its expectations are a sensible benchmark for any regulated organisation handling client or patient data.
How Rhentech can help
We provide independent penetration testing and continuous exposure monitoring through our managed IT and cybersecurity service, and we help boards define the containment and recovery decision rights the PA describes as part of incident response planning.
If you want to understand how your testing and patching cadence compares with these expectations, book a free initial consultation with a senior consultant. The consultation costs nothing; testing and managed services are paid engagements scoped to your environment.
Sources
- Prudential Authority, Communication to the Financial Sector: Preparing for Frontier AI and AI-Accelerated Cyber Risk, 24 April 2026.
- FSCA and Prudential Authority, Joint Standard 2 of 2024, Cybersecurity and Cyber Resilience Requirements, published 17 May 2024 and effective 1 June 2025.
- Minister of Finance, reply to written question NW385 on the Land Bank ransomware incident, 4 March 2026.
This article is for informational purposes. It does not constitute legal or regulatory advice. Consult qualified advisers for matters specific to your organisation.
Reviewed 20 October 2026.