AI is eroding the accidental security of being too boring to hack

For many organisations, being overlooked has quietly passed for being secure. A small business, an unremarkable supplier or a professional firm with no famous brand may have assumed that attackers had better targets. That assumption was never a technical control. It was a bet that the labour required to find, enter and monetise the organisation would exceed the likely return.

Artificial intelligence threatens that accidental shield. It does not make every company intrinsically valuable, nor does it conjure exploitable weaknesses from nothing. Its more important effect may be economic: reducing the time, expertise and repetitive work required to investigate many potential targets and process the useful ones.

This is best understood as the erosion of security by unprofitability. If the cost of attacking an organisation falls faster than the value available from it, a target that once looked marginal can become viable. Boards should therefore stop treating obscurity as part of their protection and start examining attacker effort alongside asset value, identity exposure, supplier access and recovery capacity.

The denominator is changing

Anthropic’s September 2026 threat intelligence report describes activity observed through its services between December 2025 and August 2026. The company says AI assistance can increase the speed, scale and depth of cyber operations, with the greater risk coming from support across an attack chain rather than exploit creation alone. Its case studies include reported use of direct execution and multi-agent orchestration for reconnaissance, exploitation and data exfiltration.

The consequential point is not that an autonomous machine has become an infallible hacker. Anthropic itself says humans still make important decisions, including selecting targets, reviewing results and choosing how to monetise an operation. Instead, its assessment is that lower skill and labour requirements can reduce campaign costs, support higher-volume activity and make previously marginal targets more viable.

That is a provider’s assessment based on its own visibility, not a census of cybercrime. Anthropic cannot measure malicious activity using every other commercial service, open systems or no AI at all. A separate company analysis of 832 accounts banned for malicious cyber activity found that the share rated medium risk or higher under Anthropic’s method rose from 33 per cent in the first half of the study period to 56 per cent in the second. That says something useful about the activity Anthropic detected, but not the prevalence of AI-assisted attacks across the whole economy.

Independent testing nevertheless supports taking the capability trend seriously. The UK AI Security Institute has reported rapid improvement in frontier models, including about 50 per cent average success on apprentice-level cyber tasks and the first successes on some expert-level tasks in 2025. It also found that models continued to struggle with realistic multi-stage cyber ranges. The evidence points to meaningful assistance and uneven autonomy, not effortless compromise.

Small does not mean worthless

The distinction between target value and attack cost matters. A modest organisation may have limited cash and little public profile, yet still hold credentials, customer information or access to another organisation. In Anthropic’s reported intrusion cases, most initial access began with compromised credentials. Its report also describes the discovery of exposed access tokens across applications, code, containers and related stores.

AI did not create those credentials or exposures. It may, however, help an operator search, interpret and act on them more quickly. Australian Signals Directorate guidance published in July 2026 similarly says attackers may use AI to discover and exploit internet-facing weaknesses faster, pursue many victims at once and shorten the path from initial compromise to theft or extortion.

A conditional scenario follows. If capable cyber tools continue to become accessible, if attackers can still obtain initial access through existing weaknesses, and if stolen data, access or disruption remain monetisable, then long-tail targeting could expand. Lower labour per target would let an operator test more marginal organisations. No public population-level dataset in the research establishes that this expansion has already happened, so it should be treated as a risk model rather than a measured outcome.

The same logic applies to suppliers. A small provider might be uninteresting on the strength of its own balance sheet but valuable because it holds customer data or privileged access. If attackers can identify such relationships cheaply, the supplier’s position in a network may matter more than its size. That is a possibility, not a universal description of small business, and Anthropic has not published a representative breakdown of victims by size or sector.

Australia already has little room for complacency

Australian figures do not isolate AI-enabled incidents, but they show the existing exposure into which cheaper attack labour could arrive. The Australian Bureau of Statistics reported that 21 per cent of businesses experienced a cyber security incident in 2024-25, while 28 per cent had no measures in place to prevent incidents.

ASD recorded more than 84,700 cybercrime reports in 2024-25. The average self-reported cost was $80,850 per business report, including about $56,600 for small businesses. Its annual report also cited survey findings that 22 per cent of SME owners said their business had been affected by cybercrime in 2024.

The Office of the Australian Information Commissioner received 1,205 notifiable data breach reports in 2025, the highest annual total since the notification scheme began. Of those, 716 were attributed to malicious or criminal activity, with cyber hacking remaining the primary cause. None of these figures proves that AI caused the incidents. They establish a more basic point: many organisations are already exposed, and the consequences are not confined to prominent enterprises.

That should guard against the wrong response. The lesson is not to buy an AI-labelled security product and declare the problem solved. A board cannot automate its way out of excessive privileges, exposed access, weak preventive measures or inadequate recovery planning. The relevant question is whether technology changes the time and effort needed for an attacker to exploit those conditions.

Defenders can change the economics too

There is a serious counterargument: AI can reduce defensive labour as well as offensive labour. ASD says it can assist with threat-intelligence analysis, alert triage, policy compliance and vulnerability discovery. If organisations deploy such capabilities with appropriate oversight, can act quickly on findings and can afford the tools, automation might identify exposures earlier and shorten response times.

The available evidence does not reveal which side will gain more, or whether those gains will be evenly distributed. Larger organisations may be better placed to turn faster analysis into faster remediation, while a smaller firm could remain constrained by people, money or its service arrangements. Conversely, useful defensive automation could give smaller organisations capabilities they previously lacked. The balance is not predetermined.

This uncertainty strengthens the case for an economic risk model. Boards should ask how much effort it takes to identify their exposed systems, what access a compromised identity provides, how portable their data is, whether a supplier account opens a path elsewhere and how quickly essential operations can recover. Managed service providers should be judged partly on whether they reduce those exposures and response delays, not merely on the presence of AI in a product description.

Policymakers face a similar task. Support for smaller organisations should reflect the possibility that low attacker costs can spread attention across a wider pool of targets. But policy should not be built on claims that autonomous hacking is universally reliable or that every business is now equally attractive. Capability benchmarks, provider telemetry and national incident statistics measure different things and should not be collapsed into one dramatic forecast.

Being small, quiet or boring may still affect an attacker’s calculation. It is simply not a security control. The useful response to AI-assisted attacks is neither panic nor dismissal, but a deliberate effort to raise attacker costs and reduce reusable value. Organisations cannot choose whether offensive technology improves. They can make sure that a lower profile is no longer the most important barrier standing between an attacker and their systems.

Attacker effort, identity reach and recovery capacity provide a practical lens for challenging assumptions tied to organisational profile. Build the supporting knowledge with The Hack Academy’s online cybersecurity training programme: study cybersecurity fundamentals, networking and internet vulnerabilities at your own pace, with practical exercises alongside theory.