Google’s recent confirmation that its Gemini AI hacked the systems of three companies during a cybersecurity test that was conducted in May by the security firm Irregular. The test was meant to involve simulated targets, according to the Wall Street Journal .

Gemini accessed one company’s protected system by guessing its passwords. In two other cases, it used credentials that it found in public repositories. Google said the model stopped each hack when it recognized that it had accessed a real company’s system, the Journal reported.

The incident raises an important question for business leaders: If an AI-related event affected their organization, would their crisis management plan help them respond quickly and credibly? AI tools can create new risks, but they can also help companies find weaknesses in their plans and prepare responses.

AI-related incidents such as these can cause different crisis situations for companies and organizations that can impact their operations, credibility, and reputation. The good news is that AI can also help organizations find weaknesses in their crisis management plans before someone—or something—else does.

Companies that take the time now to use AI models to test their plans will be better prepared than organizations that let their plans gather dust. The technology can catch blind spots and gaps that others may miss by running a series of what-if scenarios, including AI hacks and other technology-related hypotheticals. When a crisis strikes a company, journalists, customers, and others may turn to the internet and chatbots for information about the situation, even before they’ll visit the company’s website. That’s why companies should monitor and track what’s being said about their crisis on social media and other platforms and respond as needed.

When Jason Mudd, CEO of Axia Public Relations, ran an AI simulation against a client’s crisis plan, “it surfaced the one question their team had never asked: what happens if the spokesperson becomes the story? Another AI test flagged a full contact list for the crisis team, but no one had verified those people would actually take a call after midnight if their phones are on do not disturb mode. A third AI review caught something simpler: the plan hadn’t been touched since social media and AI itself became crisis channels in their own right,” Mudd recalled in an email interview.

The examples point to practical gaps that could delay a response or make a crisis worse. AI can help identify them, but people still need to test whether the proposed solutions work.

Asking and phrasing the right questions

Business leaders should never assume that they will never have to deal with more than one crisis at a time. That’s why it makes sense for some of the scenarios that are presented to AI models should include multiple worst-case situations. For example, how should a company respond if it is the victim of a cyberattack that shuts down its computers and communication channel, followed by allegations of wrongdoing by corporate officials?

How you ask and phrase questions to AI models will affect the answers they provide. Instead of seeking validation for a crisis plan, ask the technology to be a harsh critic and what it would take to break the plan or make it useless.

After removing sensitive or confidential information from plans, it’s time to ask AI several key questions. Scott Keever, CEO of Reputation Pros, said in an email message that the queries could include the following:

  • Where would this fail in the first 24 hours?
  • Which stakeholder is missing?
  • What would a reporter ask that we are not ready to answer?
  • Where does this statement sound defensive, vague, or legally sanitized?

Keever said the answers to the questions can expose holes in plans before the public does. It is also important to test crisis management plans against different audiences and stakeholders. “My advice is to use AI to pressure-test the plan from five angles: customer, employee, journalist, regulator, and critic. If the plan cannot survive those five audiences in a simulation, it probably will not survive the real crisis,” Keever predicted.

There should be limitations that companies place on the extent and nature of information that is shared with AI models. And don’t assume that the answers you get are completely relevant and appropriate. It’s always a best practice to have one or more humans review and evaluate AI’s recommendations before implementing them.

Business executives and their staff should exercise caution and restraint when testing crisis management plans against public AI platforms. That’s because the plans can contain sensitive operational and personnel information. “I would not put one into a public AI tool or treat an AI response as fact. AI can help a team ask more questions faster. Experienced people still have to decide what is accurate, appropriate and right for the moment,” Catherine Montgomery, a reputation and crisis management expert and founder and CEO of Better Together Agency, advised via email.

Learning from the mistakes of others

AI should not replace the role of humans in preparing for and managing crisis situations. “I’ve seen plans score clean in an AI audit, then watched a live tabletop expose the real flaw in 20 minutes: an approval chain requiring three executives who had never responded to anything in under four hours, which means 12 hours of silence during the worst 12 hours of the company’s life,” Kevin Dinino, president of KCDR PR, told me in an email message.

Companies can learn from that experience and avoid making the same mistake by:

  • Testing how they would respond to a crisis within the first 24 hours.
  • Confirming who could be reached when a crisis strikes.
  • Knowing ahead of time who would be available to approve statements about the situation.
  • Confirming how members of their crisis management teams would communicate if their usual communication channels went offline.

Reviews of corporate crisis plans by AI platforms should be followed by live tabletop exercises where participants are asked to respond to different crisis scenarios. Then, based on what the reviews and exercises reveal, revise the plans accordingly. Repeat the reviews and exercises when events warrant, such as the departure or arrival of corporate leaders, the use of new or expanded communication channels, and news or developments that could threaten the organization.

The hacking incident involving Google’s Gemini underscores the reality that AI can introduce risks that crisis plans need to address. Business executives should ensure that AI is used carefully and strategically to help strengthen crisis management plans. And don’t assume that just because the technology continues to improve, that it is perfect or will always get it right the first time.