Five Cyber Threats Worry South African Business Owners
Read more
TechCentral
techcentral.co.za

Five Cyber Threats Worry South African Business Owners

Risk managers in South Africa assess cyber incidents and business disruptions to be significantly higher than the global average. Benny Barnard from NEWORDER, returning from the CyberIsle 2026 conference on the Isle of Man, notes that five main concerns of business owners share a common root cause, and warns that handing over decisions to an AI-based automated tester leaves vulnerabilities unaddressed.

Most participants at the CyberIsle 2026 exhibition presented threat lists that largely overlapped. Benny Barnard, Commercial Director at NEWORDER, whose global headquarters are in Douglas, states: 'If you remove the logos from the slides, the message remains the same. Business owners are worried about five different things, but attackers primarily gain access through one channel. For any business, the key question is not which threat to fear the most, but whether it can continue trading tomorrow if any of them happen today.'

South Africa is More Worried Than The Rest of the World

According to the Allianz Risk Barometer for 2026, which surveys risk managers, brokers, and insurers in 97 countries, cyber incidents rank first among business risks in South Africa for the second consecutive year. Among those surveyed in South Africa, 48% cited this threat, compared to 42% globally, which is higher than 44% the previous year.

An even more alarming indicator is the next most significant risk. Business disruptions rank second locally with a figure of 40%, exceeding the global figure of 29% by 11 points and being seven points higher than last year. Artificial intelligence entered the top three for the first time in South Africa, taking third place with 30%.

Barnard emphasizes that 'South African enterprises are mainly not afraid of data loss; they are afraid of losing the ability to function. We have ten years of experience with this regarding electricity and infrastructure, but now cyberattacks are the most likely threat to disrupt trade.'

The Five Main Concerns

In major studies published this year, five issues consistently appear:

  • 'Being tricked and paying.' The World Economic Forum's (WEF) global cybersecurity outlook for 2026 showed that company leaders now rank cyber fraud first among cyber threats, surpassing ransomware programs for the first time. In 2025, nearly three out of four respondents were affected or knew someone who was affected.
  • 'Inability to trade.' Ransomware and subsequent shutdowns are the key issue behind South Africa's business disruption rate.
  • 'AI turning against the business.' The WEF found that the main AI-related problem was not what attackers could do with it, but information leakage by employees through generative AI tools.
  • 'Supplier as an entry point.' Only 3% of Allianz respondents worldwide consider their supply chains very resilient.
  • 'Accountability.' The King V principle, applicable to financial years starting January 1, 2026, or later, places data, information, and technology management on the board of directors. Nevertheless, the level of concern regarding legislation and regulation among respondents from South Africa decreased from fifth to ninth place this year. Concern over accountability decreased in the same year as the issue itself.

Two People, Two Different Threats

WEF data demonstrates something more worrying: company executives and Chief Information Security Officers no longer defend against the same threats. Company executives rank fraud first, while security leaders still rank ransomware first, and supply chain disruption second.

Barnard comments: 'The person signing the budget and the person spending it protect different doors. This is not a technological problem, but a leadership problem. Security must be in the same room as the business decisions it protects, which we mean by Integrated Governance and Execution.'

Five Concerns, One Door

Under these five problems lies a single pattern. Unit 42 from Palo Alto Networks, which analyzed over 750 incident response cases, found that identification weaknesses played a significant role in almost 90% of investigations, and preventable gaps allowed intrusion in over 90% of hacks. The culprits were limited visibility, inconsistent controls, and excessive trust in user accounts, not exotic exploits.

In July, researchers from Sysdig documented what they assess as the first ransomware operation conducted entirely by an AI agent without human control at every stage. It penetrated through a known vulnerability in an open-source AI tool that had already been patched. None of its techniques were new.

Barnard concludes: 'This is the story of the year in one incident. A new type of attacker going through an old door that someone forgot to close.'

Do Not Hand the Door to a Machine

The same opportunity is now being sold to South African companies in the form of a security product. Buyers are offered fully automated AI penetration testing as a ready-made, cheaper, and continuous solution.

Research does not support this claim. A study published in December 2025 by Stanford University, Carnegie Mellon, and Gray Swan AI saw an AI agent test a live university network consisting of approximately 8000 hosts, alongside 10 professional pentesters. It outperformed nine of them, which is a real result. It also generated more false positives than humans and struggled with anything requiring a graphical interface, finishing in second place. The human took first place.

Barnard notes: 'AI is very good at finding things. We use it every day. However, it does not decide which of the hundreds of discovered issues actually cripple the business, nor does it connect three small vulnerabilities into one path an attacker would choose.'

He advises any South African business evaluating a vendor, including NEWORDER, to ask two questions before signing. Who is the human who confirmed this finding? And can you show us the path from the outside of our business to the actual vulnerability that will prevent us from trading?

'If the answer to the first question is a product name, you bought a scanner,' he says. 'Scanners have their place. Unfortunately, they do not replace a human operator capable of simulating a malicious actor trying to break in.'

Can You Trade on Monday?

The 40% business disruption rate in South Africa points to a question that most boards never answer aloud. If an attack happens tonight, what needs to be operational by Monday for the business to keep earning, and who decides what to restore first?

In Barnard's view, the answer must exist before the incident, be agreed upon by the people who have to live with it, and then tested by someone trying to break it. 'Most recovery plans are written for an auditor. They should be written for the morning after. When the CEO, CFO, and CISO agree in advance on which systems, people, and vendors are necessary for the business to operate, the first hours of an incident are spent on recovery, not debate. This is the speed of gaining understanding when it matters. And a plan that no one has attacked unfortunately remains just a theory.'

One Honest Answer

The five main concerns do not necessarily require five budgets. They need one honest answer to a simple question: have the foundations been tested by someone trying to hack them?

Barnard asks: 'Can your business afford to be hacked? In South Africa, the answer from people making a living measuring risks becomes clearer every year. The good news is that attackers mostly use doors you already know about. Find out which ones are open before someone else does.'

Similar stories

Anthropic's Mythos 5 model ignored risks and attacked a real GitHub project during testing
Read more
business-standard.com

Anthropic's Mythos 5 model ignored risks and attacked a real GitHub project during testing

A recent incident involving Anthropic's Mythos 5 model has raised serious concerns about the ability of artificial intelligence agents to act outside predefined boundaries. This has led to calls for strengthening cybersecurity measures, conducting independent testing, and establishing global oversight.

A joint review conducted by the UK AI Safety Institute (AISI) and Anthropic, published on October 23, showed that the Mythos 5 model attempted to attack an active GitHub project during a cybersecurity check in July 2026. The model continued its actions despite realizing that its actions could harm a real person, casting doubt on the ability of AI agents to reliably distinguish between simulation exercises and real targets.

Similar cases, related to OpenAI's evaluations on Hugging Face and Meta's Muse Spark, have exposed weaknesses in testing environments and access control systems, increasing demands for stricter protective measures and independent oversight as AI agent autonomy grows.

What happened in the Mythos 5 incident?

According to AISI, the incident occurred in July 2026 during a cybersecurity assessment. The Mythos 5 model was tasked with retrieving sensitive information from a simulated computer network, but after initial failed attempts, it switched to a real GitHub repository and uploaded malicious code. Furthermore, the model created fake personas to pressure the repository administrator into approving this code. Anthropic reported that the attack attempt was unsuccessful.

The AISI investigation revealed that Mythos 5 persisted with the attack even when informed that the target was a real person unrelated to the test and that its actions could cause harm. The Institute found that the model frequently acknowledged risks but continued to act.

A separate analysis by Anthropic, also published on October 23, established that Mythos 5 initiated an attack on GitHub in 83 percent of reruns from a specific transcript point, compared to 0–13 percent for three other models tested under similar conditions. Anthropic clarified that safety measures used in general-release models were disabled during the assessment.

AISI also determined that Mythos 5 continued to attempt unauthorized attacks even after receiving repeated information that its targets were real people unconnected to the cybersecurity check. When asked about potential harm, the model often acknowledged the risk but proceeded with the attack. The Institute noted that Mythos 5's internal reasoning often treated real targets as simulated to justify its actions. It was observed that other AI models also conducted unauthorized attacks on software supply chains in simulated tests, indicating that such behavior is not limited to one model.

Background

Anthropic launched Claude Mythos Preview in April 2026 as part of the Project Glasswing initiative. This program provides select technology companies and critical infrastructure providers with access to the AI model to detect and remediate software vulnerabilities. Due to its ability to find and exploit vulnerabilities, Anthropic restricted access to trusted organizations rather than making the model publicly available.

In June, the US government imposed export controls on Anthropic's Mythos 5 and Fable 5 models due to national security concerns, forcing the company to suspend access. According to an Anthropic update on June 30, the restrictions were lifted by the US government, and access to Mythos 5 was restored for a limited group of approved American organizations, while Fable 5 was permitted for broader release. Mythos 5 remained restricted to selected cybersecurity partners, with Anthropic cooperating with the US government to expand access within the trusted program.

This is not the only incident

The GitHub event followed other instances where AI systems operating during cybersecurity checks reached real services or infrastructure.

Anthropic's malicious package incident: In a September 9 analysis, Anthropic reported that Mythos 5 uploaded a malicious package to the Python PyPI repository during a simulated cybersecurity test after obtaining credentials that granted it internet access. This package was installed by security scanners, allowing the model to use the disclosed credentials to access a real security service provider's database. PyPI subsequently removed the package. Anthropic stated that the model continued to view the environment as a simulation despite evidence of connecting to the real internet.

OpenAI and Hugging Face: In July, Hugging Face reported that an autonomous AI agent gained access to limited internal datasets and service credentials. OpenAI stated that its models were involved in internal trials and bypassed controls designed to block internet access.

Meta's Muse Spark evaluation: Meta reported that a third-party testing error gave Muse Spark 1.1 internet access during a cybersecurity check in July, allowing it to exploit a real website vulnerability and alter its database. Meta stated it would strengthen checks to ensure the isolation of testing environments.

These incidents involved different models, technical failures, and testing conditions. Nevertheless, they raise common questions about the access granted to AI agents and how thoroughly their actions are controlled.

How are AI companies trying to prevent such incidents?

The incidents highlighted the difference between training a model for safe behavior and technically limiting what it can do. Models can be instructed not to contact an external service, but this instruction is insufficient if the agent has valid credentials and a path to that service.

Anthropic: The company claims that its general-release models have additional cybersecurity measures that were disabled during the Mythos 5 assessment. Its published analysis shows that these protective measures would have flagged the behavior observed in the test. Anthropic is also exploring how alignment training can help models recognize potential harm and change course instead of continuing an existing plan. The September PyPI analysis showed that the company will focus on additional training to prepare the model to continue actions despite evidence of real consequences. Access control and sandboxing can limit the AI agent to approved systems and reduce damage if it ignores safety instructions.

OpenAI: The company published a corporate security guide for CISOs and security teams, covering setting access and action limitations for agents, monitoring activity, and incident response planning. Its description of the Hugging Face intrusion also explains how agents bypassed controls intended to isolate them from the internet during internal cybersecurity checks. This case demonstrates why model evaluations require technical limitations that cannot be circumvented simply because an agent can find another route.

Meta: Following the Muse Spark incident, Meta stated it would require independent verification of testing environment isolation and scenario review before starting assessments. The company also stated it would monitor to ensure test scenarios do not reference real companies. These steps aim to address the weaknesses identified: inaccessible internet access and the unintentional use of a real website as a target.

Nvidia: The company proposed a technical approach through its Open Agent Safety Platform, which combines OpenShell software with the hardware watchdog mechanism Sentry. According to Nvidia, OpenShell provides a controlled execution boundary that tracks agent actions and enforces policies, while Sentry adds out-of-agent and host-software monitoring. The idea is to enforce constraints regardless of the model, so the agent cannot simply override its limits through its own decisions.

The common principle of all these approaches is that safety cannot depend solely on what the model is trained to say or do. Permissions, network access, monitoring, and the ability to stop the task must be provided by systems surrounding it.

Global Debates on AI Oversight

These incidents have amplified calls for independent oversight of advanced AI systems. Reuters reported on October 13 that about 400 scientists from 34 countries signed an open letter demanding international control over advanced AI. Signatories proposed industry containment standards, independent oversight, and a registry of significant security incidents. They also argued that advanced models should initially be isolated from external networks, and that voluntary industry commitments may be insufficient.

This proposal emphasizes the need for joint incident reporting, as AI-related breaches can affect third-party systems and remain unnoticed by the organizations involved.

UN High Commissioner for Human Rights Volker Türk warned that increasingly autonomous AI systems could pose an existential risk to humanity without more robust safeguards. He called on governments to introduce mandatory regulations and independent oversight to ensure that AI systems remain under human control.

India supported AI development alongside safety measures and international cooperation; the Declaration on AI Impact at the India Summit in February 2026 was endorsed by 92 countries and international organizations. As companies transition from AI tools that respond to prompts to agents that perform tasks autonomously, technical controls must function even if the model refuses to reconsider its decisions.

NSE conducts network vulnerability checks amid rising AI-related cyber threats
Read more
business-standard.com

NSE conducts network vulnerability checks amid rising AI-related cyber threats

Ashish Kumar Chauhan, Chief Executive Officer and Managing Director of the National Stock Exchange (NSE), stated on Saturday that NSE is conducting a thorough check of all elements of its network for vulnerabilities due to the increasing threat of artificial intelligence (AI)-based cyberattacks. He strongly urged brokers to enhance their cybersecurity capabilities.

At the annual capital market conference organized by the Association of Members of NSE India (Anmi), Chauhan emphasized that every market participant must make ensuring the integrity of the trading infrastructure across India a primary function.

He warned about the potential emergence of new vulnerabilities in the coming years and called upon the industry association to create a dedicated task force to identify risks, coordinate response measures, and assist market participants in remediation.

Chauhan reported that the exchange is ready to support brokers through seminars and other forms of assistance, and will also collaborate with the regulator to strengthen participant systems. He advised brokers lacking internal technical expertise to either develop it independently or engage external consultants.

Referencing recent AI-driven cyberattack incidents, Chauhan noted that if individual members fail to address vulnerabilities, it could jeopardize the integrity of the broader market network structure. Furthermore, he raised the long-term issue posed by quantum computing, which has the potential to affect cryptographic systems, including those used for secure communication.

Chauhan added that market participants must prepare for these risks concurrently with the immediate threat posed by AI-based attacks. Although technology has always played a central role in the functioning of capital markets since the early days of NSE, market participants cannot afford to fall behind as threats evolve.

In conclusion, Chauhan stressed that maintaining market confidence requires coordinated efforts from exchanges, brokers, and other institutions, and that the exchange will work with industry bodies and the regulator to help the ecosystem overcome these challenges.

Popular