AI Can Keep Probing While a Hacker Walks Away. That Changes the Workload.

Small online retailer’s packing desk with a conceptual shop and security display.
AI-generated illustration of protecting an online shop; not a victim’s website or an actual attack dashboard.


The detail that caught my attention in this AI-assisted hacking investigation was the amount of work a person could delegate. A short instruction could set software probing for weaknesses over many hours, across many targets.

For a small business or an independent website, that is a more practical concern than imagining an all-powerful machine. The question is whether ordinary weaknesses become easier to pursue at scale.

What was observed, and what remains uncertain

Reuters’ investigation reports more than 1,200 attempted website intrusions and over 100 compromises associated with one Chinese-speaking operator. Instructions were under 100 Chinese characters. Reuters could not establish how much data, if any, had been extracted without authorisation.

The accompanying NorForge research describes exposed control panels for Cairn, a general-purpose multi-agent framework. Its snapshot identifies over 1,100 attack projects in the most active cluster, plus other clusters. That project count and Reuters’ later site count should not be treated as identical measurements.

NorForge reports checking some claimed compromises directly. However, agent-reported success is not automatically an independently confirmed breach, and the objectives differed across projects. Language also does not establish an operator’s nationality or state affiliation.

Cheap persistence changes the workload

For successful projects in its most active cluster, NorForge observed an average time to goal of just under 23 hours. It estimated roughly $5 per task including unsuccessful attempts, with considerable uncertainty and some much more expensive runs. It did not have actual token-usage totals.

The same caution applies to the approximately $5,000 campaign estimate reported by Reuters: it is a researcher’s estimate, not an audited invoice.

I would not turn either figure into a universal price for breaking into a website. What matters is the possibility of sustained work without matching human attention minute for minute.

An unsuccessful first attempt may no longer be the point where an attacker moves on. Software can keep exploring alternatives, while the person directing it starts another task. That makes neglected systems and incomplete inventories more worrying.

AI does not make the basics irrelevant

The UK National Cyber Security Centre’s assessment expects AI to make parts of cyber operations more effective and accessible. I read that as a reason to shorten the gap between discovering a weakness and dealing with it, rather than abandon established protection in favour of a fashionable new product.

For a small operation, the NCSC’s practical guidance is a useful starting point. My priorities would be:

  • Know what is online. Include old subdomains, support portals and third-party services, not just the main site.
  • Keep supported software current. Remove services and plugins that no longer have a purpose or a maintenance plan.
  • Protect administration. Use strong authentication, restrict access and remove accounts that no longer need it.
  • Make recovery real. Keep backups separate and check that restoring them works.
  • Give alerts an owner. Logs only help when somebody can recognise a problem and respond.

On a managed platform such as Blogger, some infrastructure responsibilities sit with the provider. My own account security, collaborators, added scripts and domain settings still deserve attention. There is little value in buying a tool to manage servers I do not operate while neglecting the account that controls the site.

More urgency, without giving up perspective

I covered the difficulty of attributing AI involvement in my post on South Korea’s bank breaches. This investigation offers a closer look at agent workflows, but it still has visibility limits.

The sensible response is to make routine protection more dependable. I do not need to know whether the next attempt comes from a person or an agent to want fewer exposed weaknesses and a workable recovery plan.

Comments

Popular posts from this blog

Obsidian’s New Search: Less Friction for Notes in Two Languages

OPPO Find X9 Ultra vs Find X10 Pro Max: Is the Older Ultra Still Worth Buying?

OPPO Find X10 Technical Preview: Cameras, Design and Missing Specs