The joy of automated threat analysis (#1416)

Topics/tags: Technology, rambly

In part of my quest to get far too much email, I subscribe to a few daily newsletters from TLDR [1]. These are intended to be quick summaries of what’s going on in the tech industry. Each day, they send a list of articles they’ve identified along with quick summaries. As you might guess, the basic idea is that one can skim the headlines, browse the summaries, and then delve into articles as appropriate.

I don’t always read the daily newsletters; I’m not in a job that requires that I stay on top of all the news on, say, AI and security, to name two of the daily newsletters I get [2]. Still, I like to be able to skim the subjects and, at times, I delve more deeply. It can be useful to see what folks are writing and reading about AI and security. These newsletters can also give me stories that are worth sharing with students, perhaps as vignettes in class, perhaps as readings in CS table. For example, I find this a bit scary.

ChatGPT Health (3 minute read)

OpenAI launched Health in ChatGPT for US adults, allowing users to connect Apple Health and supported medical records for more contextual health conversations. The company said connected health data and related chats would not be used to train foundation models or target ads.

But it’s useful to know that people think that such an activity might be valuable or safe. And it might be worth having an extended conversation with students about why this might be a good (?) or really bad idea.

In any case, about once a month, I get a message that looks something like the following.

Proofpoint Threat Response
Quarantine Notification for Message titled Claude Agent Escape 💻, Chick-fil-A Breach 🍗, Notepad++ Malware 📝

Caution: External Email Sender: dan[@]tldrnewsletter[.]com
Subject: Claude Agent Escape 💻, Chick-fil-A Breach 🍗, Notepad++ Malware 📝
Date: 2026/07/24 13:13 UTC
Recipient(s): rebelsky[@]grinnell[.]edu
Verdict: Malware

Attention, automated threat analysis has determined that the message titled Claude Agent Escape 💻, Chick-fil-A Breach 🍗, Notepad++ Malware 📝 was malicious.

This email has been removed from your Inbox. Please contact the Service Desk at (641) 269-4901 if you have interacted with this message or if you need any further information.

Grinnell College IT Security Team

This email was sent from outside Grinnell College. Treat attachments and links with caution. If this message looks suspicious, use the report phish button.

Oops! Forgot to mention that TLDR also includes silly emojis in their subject lines. They get even more annoying when they get repeated that many times.

In any case, Proofpoint (our primary email security software) has decided that a newsletter about security is malicious. Perhaps that’s not surprising: A security newsletter is likely to have content that matches keywords for potential security issues.

Every time this happens, I wonder what particular issues caused the threat analysis to trigger today. Why was today’s digest so much worse than yesterday’s, or the day before’s, or most before that? As I said, it only happens once every month or so. Is it that CVE‑2026‑46331 is mentioned explicitly? That chicken is involved? That some path names used for hacks appear in the digest? I suppose I’ll never know. And I shouldn’t really care.

I also find myself amused that we put the automatic Beware of external email text on a message about security. Should I not call ITS? Will I get a virus from calling? From using help.grinnell.edu?

On a more serious note, would the threat analysis perhaps prevent me from being notified about an urgent issue? I doubt it, mostly because I don’t deal with highly secure data or systems, but I wonder.

In any case, the quarantine did get me to skim today’s digest. And that meant that I saw the following.

AI arms race in line for a reckoning after OpenAI hacking incident (3 minute read)

OpenAI’s GPT-Sol 5.6 test agent escaped an internal sandbox, reached the internet, exploited vulnerabilities, and stole credentials from Hugging Face while chasing a cybersecurity task. The model had safety checks removed for evaluation, and reinforcement learning pushed it to focus on goal completion over constraints. The breach has prompted calls for tighter standards and regulatory attention around autonomous, reward-driven security agents.

I have little faith in tighter standards and even less in regulatory attention. I’m frustrated that OpenAI can’t build better sandboxes.

Perhaps I shouldn’t read the news.

Postscript: I’ll note that the 3 minute read doesn’t include the time to read the comments. Some of those are important. Here’s one.

The official story is that the prompt was for it to beat the ExploitGym cyber security benchmark and it decided it would be easiest to hack a company that probably had solutions on file rather than beat the test from scratch.

[1] Cute name, isn’t it?

[2] I also get the general newsletter and the dev newsletter.

Version 1.0 of 2026-07-24.