AI wants to please its human. Through its training, it will run a “thought” process in many directions at once until it gets to its goal.
As you would hope, there are good, strong guardrails baked into most AI platforms, but most is not all.
There are many publicly, easily downloadable models that are tweakable, tuneable and guardrail free. They will happily do the bidding of a mildly technical human who is at best curious, at worst nefarious.
Up until now, hacking has always had many barriers to entry. It takes years of experience and months of work to hack a target, if it’s vulnerable at all.
But privately hosted AI is so fast, so capable and trained on all the latest hacking techniques that it only takes a motivated person a few hours to achieve what once took six weeks.
Website defacement
AI is commanded by natural language instructions. It really is as simple as a hacker typing in “use your resources to infiltrate the website https://charity.co.uk and change the homepage to read ‘Hacked by Ghost500’”.
This means hackers using AI can deface any page to say anything the infiltrator wants, perhaps something political or divisive, giving the charity a very difficult PR problem.
Some website takeovers are done to enable fraud. This could be subtle, such as adding a new feature article titled: “We are encouraging young people to invest in crypto, start by clicking here and investing a small amount of money and gain 10x returns”. It would seem to have the endorsement of the charity.
Another defacement could be a fake warning message urging visitors to “click here” to scan for malware. Of course, the scan is fake and is in fact the malware they are trying to distribute.
There are many variations of this, too many to illustrate but they are all very much unwanted by the charity and could damage the branding and trust it’s worked so hard to build.
While public-facing AI service providers will refuse such requests, privately hosted versions have no such policing. They will, with some tweaks that are easy to find on the internet, do the normally prohibited task with enthusiasm.
Vulnerability
Before the 2020s, hacking was getting harder as systems strengthened. But AI hacking is far outpacing the work that’s necessary to harden a system.
Just last month, AI found a longstanding but hard to spot weakness in commonly used tool PHP. Websites such as the Association for Charitable Foundations’ have been taken down for hardening work rather than wait to be a victim.
In the UK and around the world, liability for autonomous agents is legally uncharted. If a server is hacked, who’s the perpetrator?
Charities can be especially vulnerable to this, with tight budgets and staff and volunteers often wearing many hats.
Protecting against attacks
How does a charity board protect the charity with the resources available? In short, it can’t – it must draw on experts and of course that needs funding.
It may be difficult but reaching out to supporters, donors and benefactors with an aim to raise the funds to harden the digital footprint of a charity is a worthy cause. A modern equivalent to raising money to replace the church roof.
But there are some major action points any charity should consider:
- Deploy modern, strong security on your website, with active monitoring and reportage to someone who’ll actually read it and react.
- Monitor your website closely. There are plenty of free ways to watch for changes.
- Be able and know how to roll back the website to its proper content quickly.
- Have a PR message written and on standby, ready to post on the recovered website and perhaps send to supporters and beneficiaries.
This just covers the website, the most commonly exposed and vulnerable part of a charity’s IT system. The back office and all its related interconnected systems are very likely vulnerable too.
With AI commanded to look for any way to compromise an organisation’s system there has never been a more demanding need for very strong IT security and procedures.
Paul Appleby is founder of Appleby Systems
