Hello fellow keepers of numbers,

Give all your friends working on IT security a hug this weekend. It’s been a crazy one.

This week, the last 6 months caught up to us. There were a number of major cybersecurity incidents that were AI-assisted. We also received word of a new type of certification for AI products and AI agents that will have a CPA stamp of approval. It involves testing AI for hallucinations, data leaks, and accuracy.

Plus, stick around for a demo showing how to use Claude to help you track your exposure to the current, and future, cybersecurity incidents.

Also, next week will be an off-week for the newsletter as I travel for a loss in the family. But I’ll be back the following week to catch up on what we missed.

The Latest

Several major cybersecurity incidents came to light this week, alongside urgent fixes for widely used software. The reports include stolen company data, exposed passwords, and AI-assisted attacks on business systems.

Amazon says AI-assisted analysis is helping attackers turn newly disclosed flaws into working attacks within hours or minutes, shortening the time businesses have to install security fixes.

Why it’s important for us:

This is outside the scope of what I typically cover in the newsletter, and undoubtedly outside the scope of my expertise. But this is probably among the most important things I could write about at the moment.

Everyone in the world should have their eyes on this. Go to your platform of choice and find someone who covers it in a way that’s easy for the layperson to understand. Search for “Black October” or “AI cyber incidents.” Or better yet, use your AI of choice to research this and help you understand.

There’s a lot of fear mongering on socials, so it’s sometimes hard to truly grasp the risks of AI at times like this. Here is my best and most realistic understanding of the cyber events this week…

Several major software companies were attacked. Sensitive data, including secure passwords and keys, were stolen, potentially giving hackers a way into more company systems. A programming language used by WordPress and many other websites also had several critical vulnerabilities that were identified, which put the sites at risk of being hacked. This could allow hackers to take over a site or leave behind malware that puts visitors at risk. Security flaws were also found in Node.js, software used to run websites and apps, potentially putting those systems at risk. And one of the largest health data companies was breached, with customer data, including PHI and PII, likely accessed by hackers.

So why is this happening all at once?

Approximately 5-6 months ago, OpenAI and Anthropic trained a class of models capable of unprecedented levels of cybersecurity. This includes the ability to find and patch software vulnerabilities that were previously unknown, many of which have existed since the software or programming language was created.

Anthropic CEO, Dario Amodei, predicted other open source models would reach the same capabilities within 6-12 months. We’re right around 6 months out from that prediction, and we’re seeing open source models perform at, or even above, the levels of the initial cyber models.

So what does this mean?

Honestly, this is where I’m still a little unclear. But it seems like bad influencers, hackers, and even adversarial countries now have access to models capable of finding and exploiting critical vulnerabilities in nearly any software or website with a prompt as simple as “figure out how to hack X.” I have no idea what the internet looks like when that’s possible or actively happening every day. But I do know it’s extremely important to follow security best practices, at a minimum.

I think we’re likely going to have to come to terms with the fact that this type of news will be a much more regular occurrence now. This is probably a new concern for many small businesses and firms, but it’s one we should all observe closely.

The AICPA and Artificial Intelligence Underwriting Company (AIUC) announced a joint assurance offering that combines a SOC 2 report with an AIUC-1 certification in one CPA-led engagement. The offering covers both an organization’s controls and the behavior of its AI systems and agents.

SOC 2 examines controls related to areas such as security, availability, confidentiality, and privacy. AIUC-1 adds testing for AI-specific risks, including hallucinations, data leakage, and attempts to manipulate an agent’s instructions. AIUC says a typical evaluation tests thousands of scenarios through repeated conversations with the agents.

The joint offering includes annual renewal and quarterly technical retesting of the AI systems under AIUC-1.

Several CPA firms are already delivering the combined offering to clients. A full launch of the joint certification is expected in early 2027, after the initial engagements conclude.

Why it’s important for us:

I'm glad to see this getting attention. AI still feels like the Wild West, and there's something fun about being able to build an agent and see what it can do. At a certain point, one person testing it one time isn’t enough. This is different from previous software features that did one thing and it worked or it didn’t. AI has a level of inconsistency, and unexpected actions can be taken without a clear reason if you’re not careful.

So testing the AI products is a great idea. Does the agent make things up? Can it leak private data? Can someone trick it into doing something it shouldn't? These are all important questions, and we’re not testing them nearly enough right now.

But I have some reservations about who does that work. I say this as a CPA and accountant (albeit, not a very good one)… I’m not sure a CPA qualifies someone to determine whether an AI agent is secure. I'd like people with serious AI and security expertise developing the methodology and performing the technical testing. A CPA-led engagement that brings in those specialists could make sense. It sounds like AIUC-1 was developed with some of these experts, which is good. But I'd like to understand more about how that responsibility is divided for this certification.

My current read of this is that a specific AI product or AI agent is what gets certified. This doesn’t appear to be a certification for all the AI products and AI agents of a particular company. This could be a certification of a new AI agent set to be released in an existing software, for example.

For a company selling an AI product to other businesses, independent testing is very reasonable. We all want to use products that are safe and secure. But I’m also a bit worried that strict guardrails on AI agents like this could slow down progress, especially in a field that requires the level of security that accounting does. This could put the accounting and finance industries materially behind the progress of other companies using AI.

I still like the direction. I just want the right expertise behind it and a process that leaves room to build useful things safely. I’d like to see more details on the testing procedures and some of the dynamics of the certification. I'd hate for an effort to make AI safer to become another reason firms and accountants stop experimenting.

Trending News

Anthropic opened a waitlist for Claude Duet, a proactive assistant that works across connected apps: This is their answer to OpenAI’s Dots, Grok Bot, Muse, and a long-running list of personal agents. It’ll probably be mostly useful for personal life early on. Once you can set up multiple agents, similar to an org chart in a business, I expect this to become much more valuable.

AICPA published a practice aid covering AI use by audit clients and audit firms: I've been pretty critical of the lack of useful guidance in the accounting space, so I'm glad to see some focus.

Xero connected live financial data to Gemini and announced client switching across its AI integrations: Xero in Gemini, cool. Not why I care. Sounds like Xero now offers the ability to add the connector and then use AI to switch between clients via the connector. This is one of the biggest pains with the integrations currently, so this would be a massive improvement.

1Password introduced broader AI agent integrations that let agents sign in without seeing passwords: With all the cybersecurity issues and data breaches this week, this feels like a timely announcement. Your passwords remain hidden from your agents at all times. I expect a lot of other password managers to deploy something similar.

OpenAI added real-time collaboration and direct editing to ChatGPT Slides: I found out about Slides last week during my Put It to Work demo, but they officially announced it this week. I’m not sure if this will immediately be as good as one of the add-ins for PowerPoint, but I can see this being much better long-term as it matures.

OpenAI highlighted shared Team Tasks for recurring work in eligible Enterprise workspaces: This was first announced at OpenAI Dev Day a few weeks ago, but is now rolling out. I expect this to make it to the Team plan in the near future as well. Shared Team Tasks make it possible to set up automations where multiple team members have access to edit them and/or view the results.

BizBuyData offered to pay accounting firms for historical records used to train AI agents: Sights have clearly been set on AI in accounting. This is a bit alarming, but selling scrubbed data happens quite often. I don’t love the idea that it’s likely being done to replace human accountants…

PwC reported $3 billion in investment across its network, particularly in AI and technology, including its PwC One platform: That’s a lot of money pouring into AI in one year.

OpenAI made its Decisions API generally available, adding caching, structured inputs, and voice integration: I've covered the Decisions API before, but it’s now generally available. This could make a big difference in accounting software and using AI to cheaply and efficiently automate accounting workflows.

Grok Bot added access to Calendar, Reminders, Contacts, and other phone apps on iOS and Android: Grok Bots that can use your phone’s apps. This actually feels pretty useful. I hate Apple Reminders, but I still use them. An AI model that does my reminders feels way better.

Put It to Work

This might be one of my more boring demos, but probably a very important one. Given the news this week about all the cyber incidents that could impact us, I thought it’d be useful to show how you can use Claude (or ChatGPT) to track the incidents for you. It identifies areas where you may have exposure and offers recommendations.

Use this to track anything you’re doing in response to a data security breach, which could be useful down the line for insurance or a number of other reasons.

If you’re feeling overwhelmed by constantly trying to track this, ask Claude or ChatGPT to turn the research and logs into a weekly scheduled task for you.

Weekly Random

Anthropic announced plans to put 5% of its shares into a new public benefit fund. One listed use would be direct payments to people affected by AI’s impact on jobs.

Just last week, I wrote about proposed taxes to fund support for workers as AI changes the economy. This feels even better because Anthropic is voluntarily (I think?) setting aside equity, so the fund’s value grows with the company’s value.

Having a connection between the fund’s value and AI’s success makes a lot of sense. If AI creates wealth, some of it should help the people who are disrupted along the way. If AI is halted for any number of reasons, the fund isn’t very relevant, but fewer people are likely displaced.

After seeing Anthropic’s announcement, Elon announced he’d give away a share of SpaceXSI (yes, they changed the name again) to every child in the United States next year. I guess that’s good? The idea is at least somewhat similar.

It’s encouraging to see even a small level of focus on this area as we move closer toward AI that will displace human workers. It’s already happening in some regard, but as models get more capable of long-term, sustained goals and memory, it’ll become even more common.