Modern Cyber · Episode
142

This Week in AI Security - 19th November 2026

A packed episode this week, with agentic attacks up front and a closing question that every security team should be able to answer.

This Week in AI Security - 19th November 2026

Podcast Transcript

Welcome back to This Week in AI Security, coming to you for the week of the nineteenth of November twenty twenty six. And we have a lot to get through this week, as with many weeks recently. So without further ado, let's go ahead and dive in on some of our top stories for this week. We've got a couple of different categories, and what I want to start off with is kind of agent or agentic attacks.

So Google finds hackers using AI agents to build and run exploits from software packages. We've talked about the apocalypse and the rapid discovery of vulnerabilities by AI tools across basically the entire software landscape. But one of the interesting things to come out of this now is that there is evidence from the Google Threat Intelligence Group that some of the patches that are being developed to patch vulnerabilities are being analyzed for further vulnerabilities that are being introduced as part of the patch, and that is creating this kind of self-reinforcing cycle. And it looks like on a lot of things where the source code is available, it's probably going to go through a little bit of a game of whack a mole in terms of different cycles of vulnerability. Patch. Patch introduces new vulnerability. New patch created. New patch creates new vulnerabilities. Etc., etc. And so on down the line until we get to a place where we're really happy and we've got very, you know, much more secure software to run. And if we put this in context as far as, like, the scale of this, just to look at zero day exploits across common third party open source software, we had twenty two of them identified in August of twenty twenty six, compared to eight a year ago, in twenty twenty five. So the scale, and really the sizing and the scope of packages that are being inspected and vulnerabilities that are being found, is up almost, you know, two and a half X in that time period. All right.

Moving on to kind of autonomous or agent powered attacks. The FBI is warning that the RTX AI hacking tool is now powering global cyber attacks. We discussed this on a previous week. We noted, by the way, that the original RTX repo was taken down on the eighth of October by the original author, but that many forks existed, and there were one thousand six hundred exposed consoles of RTX recently discovered using, I think, Shodan, that does an internet scan looking for particular known fingerprints. It's important to understand RTX is just a framework for running attacks. It is not a model of its own. It plugs into many different models, GPT models from OpenAI, Claude models from Anthropic, DeepSeek, GLM, Grok, Kimi, etc., and it is being used by attackers to target organizations, gaining initial access and then moving into privileged access within a time frame as short as three hours right now. So very powerful toolkit, lots of techniques embedded in them. None of the techniques are new. And so the key thing to understand here is really what this does is it just puts the whole process on steroids, so to speak, in terms of both the volume of attacks that can be run, but most importantly, how quickly each attack can be run on its own. All right, moving on to our next story.

Still on the theme of agentic or agent powered attacks. Seventy eight thousand character jailbreak prompt, no refusals across seventy plus sessions, is kind of a headline, but really, the net effect of it is that an AI agent ran an intrusion across a telecom provider. The operator even checked in on the environment. This was discovered by a team called Malware Labs. Kudos to them. They found a jailbroken DeepSeek Flash agent running a two week intrusion on a European telco, while its operator only checked in from time to time. The operator didn't need to know about each technique, but just checked in to make sure that the agent was continuing to infiltrate the organization. There are indicators of compromise that you can look at in terms of different traffic or activities that the agent takes as it goes through. Initial access appears to have been through CVE-2026-20320, which affects the BroadWorks platform. So if that's something that you're running in your own environment, you're going to want to patch, but it's just, again, more evidence of the fact that, you know, agent or agentic powered attacks are a reality right now. All right, moving on to our next topic for this week.

Researchers disclose a critical Java flaw with the help of AI. This is a flaw that has apparently existed within an Oracle Java product, or I guess the Java framework is probably the right way to think about it. Going back, I think more than a decade was the time frame that I saw. A pretty trivial thing: sixty four A characters plus an equals sign, right? Sixty four bytes into a buffer as small as fifty creates a buffer overflow, which triggers unexpected behaviors that can lead to, like, privilege escalation, privileged access, etc. The interesting things about this, first, by the way, Oracle patched this the same day that it was reported. So real kudos to the Oracle team for rapid response. I think that is an indicator of where software teams probably need to be, the level of responsiveness that they need to display here in twenty twenty six as some of these real critical severity bugs are disclosed to them. But the thing that I wanted to highlight is that this is a flaw that's existed for a long time. And it was found in under one hour using Codex, reportedly with under two hundred dollars worth of tokens to discover that. And this is something. Java is in use by large enterprise organizations all around the world. So think about the calculus here. Two hundred dollars worth of tokens to discover a flaw like this. We're lucky that this was security researchers who went straight to Oracle to disclose this, as opposed to threat actors who could have taken this and tried to attack any number of organizations running Java, which, by the way, probably includes your bank and definitely includes pretty much every airline on the planet, right? So think of the sensitivity of the data that is potentially accessible here. All right, moving on to our next story for this week.

Prompt injection attacks in GitHub Actions. This is actually nothing really, really new, but I just wanted to kind of continue because we've touched on this theme so many times. Anytime you hook up a software development environment to an agent or an LLM of some kind, any text can be an attack surface here, right? So in this case, this was something that was passed to OpenAI on their Codex platform, with prompt injection embedded into the issue title and body. And they were passed straight into the agent prompt and then executed from there to execute that prompt injection. OpenAI did fix it. They did pay the bug bounty researcher who disclosed it to them. This was disclosed on the eleventh of November twenty twenty six. Awesome.

So moving on to a couple of CVEs. Not much to touch on here. One is around the AWS Labs Timestream for InfluxDB MCP server. So just a couple of things to note. Again, it's on the theme of vulnerabilities being discovered at record pace here. This is CVE-2026-126681, affecting that piece of software up to version 0.0.9, fixed after that. Rotate tokens. Make sure you're using the latest version. Another CVE in Visual Studio Code, injection through MCP elicitation notifications, with patch analysis, CVE-2026-12779. It basically indicates that a malicious MCP could place a clickable workbench command inside an ordinary Visual Studio Code notification, and a single click on that link would run with the user's permission. So if you've got a privileged user or a developer with access to a bunch of systems that could have been affected, there, again, the fix is to update to the latest version. This is fixed as of VS Code 1.145.3 as of the tenth of November.

All right, moving on and changing gears a little bit, let's talk about the Astra model family from OpenAI in the GPT family of models. You might remember that GPT-6.1 Astra was canceled for scope during lab testing, cyber lab testing in particular. It showed a lot of signs of doing things that it was not asked or allowed to do. GPT-6.2 has now shipped, and some of the benchmarks that OpenAI went to, in terms of updated testing to see whether it was safe for release, was they replayed sessions where an earlier agent went wrong. So from 6.1 to 6.2, the failure rate went down. And also, importantly, in a 6.2 lab environment, none of the honeypots were triggered. And honeypots, if you're not familiar, these are basically kind of deception technology or decoys sitting within a network that might look like they expose a particular vulnerability. And it's a good way to see if a particular piece of malware is present in your environment, because if it is, it should presumably hit that vulnerability. So it's a good way to detect, hey, I've got something malicious going on here. It's just, like, basically set off an alarm bell. Also noted here, about half as often as its predecessors did it take improper actions. It is also cheaper and stronger, eight dollars, forty dollars, out of the million token cost here. So still a few kind of things to think about here from the standpoint of releasing models that have advanced capabilities that might have a little bit of danger, but kudos to them for also, you know, running some of the testing around it. And at the same time, OpenAI has taken the additional step of signing agreements with four independent evaluators to boost the company's safety work.

I think this is actually a really positive response. If you go back to some of the earliest episodes of This Week in AI Security, you might have picked up a theme around that time that Anthropic was really kind of much more leading the way in terms of providing predictability, guardrails, and cybersecurity safety controls for their models at that time. And that was one of the reasons that at that time Anthropic was perceived as being the more, quote unquote, enterprise friendly model provider, potentially. But now we're at a point where I think, like, OpenAI has really invested a lot of time, effort, and money into bridging that gap and is probably on par, at least with Anthropic. I'm not qualified to say whether they might be even ahead in terms of providing safety controls and guardrails at this point in time, but good to see that happen. They've got a number of partners now that they're working with. Some are on paid engagement, others are volunteers that are, you know, let's say, sponsored by NGOs or other organizations that are looking to promote overall AI safety around the world.

And moving on to our last story for this week. And this is around the Fireworks.ai platform. Now Fireworks.ai, in case you're not familiar with it, bills itself as a specialized intelligence platform. You can connect your data sources, some of your own infrastructure. If you've got your own GPU infrastructure, run some of your own models on that infrastructure, which basically gives you a little bit more control in terms of the data that you might connect to different LLMs and different AI systems in order to kind of move past initial general purpose queries into more specific things trained on your data sets, etc., etc. Now they have had incidents in October and in November where their platform was compromised. Some of that compromise came via something called a Cursor VS Code extension. We've talked about VS Code already in this week's episode. Cursor is an AI coding platform similar to a Lovable or a Bolt, or any number of others that are very similar. But the interesting thing to note here is that as part of this breach, transcripts of some of the chat sessions were disclosed. Some of those sessions appear to have contained pasted credentials. So that would be a developer who's building something on Cursor to connect to Fireworks AI that might have, let's say, an API key or an application token or something like that present in order to facilitate the connection either to an application or between an application server and a database server or whatever the case may be. There, like I said, was an initial breach in October. There was another one in November. Customers have been informed. All the customers who were affected have already been informed as well individually.

But here's the key thing to ask yourself: is anybody in your organization using Fireworks? Is anybody in your organization using Cursor? Would you be able to have definitive data to tell you, no, we're not using any of these platforms? Second question that goes along these lines: do you know that you don't have developers pasting credentials into any of the AI coding environments that they're working with? These are the key questions that I would leave everybody with regarding this incident and any number of other incidents that are coming. The interconnectivity of the AI ecosystem is kind of growing very, very rapidly, with specialized tools and specialized tool vendors popping up to fill lots of different niches. And it does become the most expedient way to build software, by connecting the leading tools that get you all the capabilities that you might want to move forward and innovate very, very rapidly. But that exposes you to risk. And each one of those can represent a third party risk to your organization in terms of cyber incidents that they might have, and in terms of your data that might live in those data sets when those platforms do get breached.

All right. So we will leave it there. If you are interested in understanding your own exposure, your own AI footprint, please do feel free to reach out to us and we'll be happy to answer any questions that we can, as well as help you gain visibility into any shadow AI that you have. And thank you so much for listening to This Week in AI Security. We will talk to you next week. Bye bye.

‍

Protect your AI Innovation

See how FireTail can help you to discover AI & shadow AI use, analyze what data is being sent out and check for data leaks & compliance. Request a demo today.