This Week In Security: The AI Hacker, FortMajeure, And Project Zero

August 15, 2025 by 5 Comments

One of the hot topics currently is using LLMs for security research. Poor quality reports written by LLMs have become the bane of vulnerability disclosure programs. But there is an equally interesting effort going on to put LLMs to work doing actually useful research. One such story is [Romy Haik] at ULTRARED, trying to build an AI Hacker. This isn’t an over-eager newbie naively asking an AI to find vulnerabilities, [Romy] knows what he’s doing. We know this because he tells us plainly that the LLM-driven hacker failed spectacularly.

The plan was to build a multi-LLM orchestra, with a single AI sitting at the top that maintains state through the entire process. Multiple LLMs sit below that one, deciding what to do next, exactly how to approach the problem, and actually generating commands for those tools. Then yet another AI takes the output and figures out if the attack was successful. The tooling was assembled, and [Romy] set it loose on a few intentionally vulnerable VMs.

As we hinted at up above, the results were fascinating but dismal. This LLM successfully found one Remote Code Execution (RCE), one SQL injection, and three Cross-Site Scripting (XSS) flaws. This whole post is sort of sneakily an advertisement for ULTRARED’s actual automated scanner, that uses more conventional methods for scanning for vulnerabilities. But it’s a useful comparison, and it found nearly 100 vulnerabilities among the collection of targets.

The AI did what you’d expect, finding plenty of false positives. Ask an AI to describe a vulnerability, and it will glad do so — no real vulnerability required. But the real problem was the multitude of times that the AI stack did demonstrate a problem, and failed to realize it. [Romy] has thoughts on why this attempt failed, and two points stand out. The first is that while the LLM can be creative in making attacks, it’s really terrible at accurately analyzing the results. The second observation is one of the most important observations to keep in mind regarding today’s AIs. It doesn’t actually want to find a vulnerability. One of the marks of security researchers is the near obsession they have with finding a great score. Continue reading “This Week In Security: The AI Hacker, FortMajeure, And Project Zero” →

This Week In Security: Trains, Fake Homebrew, And AI Auto-Hacking

July 18, 2025 by 7 Comments

There’s a train vulnerability making the rounds this week. The research comes from [midwestneil], who first discovered an issue way back in 2012, and tried to raise the alarm.

To understand the problem, we have to first talk about the caboose. The caboose was the last car in the train, served as an office for the conductor, and station for train workers to work out of while tending to the train and watching for problems. Two more important details about the caboose, is that it carried the lighted markers to indicate the end of the train, and was part of the train’s breaking system. In the US, in the 1980s, the caboose was phased out, and replaced with automated End Of Train (EOT) devices.

These devices were used to wirelessly monitor the train’s air brake system, control the Flashing Rear End Device (FRED), and even trigger the brakes in an emergency. Now here’s the security element. How did the cryptography on that wireless signal work in the 1980s? And has it been updated since then?

The only “cryptography” at play in the FRED system is a BCH checksum, which is not an encryption or authentication tool, but an error correction algorithm. And even though another researcher discovered this issue and reported it as far back as 2005, the systems are still using 1980s era wireless systems. Now that CISA and various news outlets have picked on the vulnerability, the Association of American Railroads are finally acknowledging it and beginning to work on upgrading.

Putting GitHub Secrets to Work

We’ve covered GitHub secret mining several times in this column in the past. This week we cover research from GitGuardian and Synacktiv, discovering how to put one specific leaked secret to use. The target here is Laravel, an Open Source PHP framework. Laravel is genuinely impressive, and sites built with this tool use an internal APP_KEY to encrypt things like cookies, session keys, and password reset tokens.

Laravel provides the encrypt() and decrypt() functions to make that process easy. The decrypt() function even does the deserialization automatically. … You may be able to see where this is going. If an attacker has the APP_KEY, and can convince a Laravel site to decrypt arbitrary data, there is likely a way to trigger remote code execution through a deserialization attack, particularly if the backend isn’t fully up to date.

So how bad is the issue? By pulling from their records of GitHub, GitGuardian found 10,000 APP_KEYs. 1,300 of which also included URLs, and 400 of those could actually be validated as still in use. The lesson here is once again, when you accidentally push a secret to Github (or anywhere on the public Internet), you must rotate that secret. Just force pushing over your mistake is not enough.

Fake Homebrew

There’s a case to be made that browsers should be blocking advertisements simply for mitigating the security risk that comes along with ads on the web. Case in point is the fake Homebrew install malware. This write-up comes from the security team at Deriv, where a MacOS device triggered the security alarms. The investigation revealed that an employee was trying to install Homebrew, searched for the instructions, and clicked on a sponsored result in the search engine. This led to a legitimate looking GitHub project containing only a readme with a single command to automatically install Homebrew.

The command downloads and runs a script that does indeed install Homebrew. It also prompts for and saves the user’s password, and drops a malware loader. This story has a happy ending, with the company’s security software catching the malware right away. This is yet another example of why it’s foolhardy to run commands from the Internet without knowing exactly what they do. Not to mention, this is exactly the scenario that led to the creation of Workbrew.

SQL Injection

Yes, it’s 2025, and we’re still covering SQL injections. This vulnerability in Fortinet’s Fortiweb Fabric Connector was discovered independently by [0x_shaq] and the folks at WatchTowr. The flaw here is the get_fabric_user_by_token() function, which regrettably appends the given token directly to a SQL query. Hence the Proof of Concept:

GET /api/fabric/device/status HTTP/1.1
Host: 192.168.10.144
Authorization: Bearer 123'//or//'x'='x

And if the simple injection wasn’t enough, the watchTowr write-up manages a direct Remote Code Execution (RCE) from an unauthenticated user, via a SQL query containing an os.system() call. And since MySQL runs as root on these systems, that’s pretty much everything one could ask for.

AI guided AI attacks

The most intriguing story from this week is from [Golan Yosef], describing a vibe-researching session with the Claude LLM. The setup is a Gmail account and the Gmail MCP server to feed spammy emails into Claude desktop, and the Shell MCP server installed on that machine. The goal is to convince Claude to take some malicious action in response to an incoming, unsolicited email. The first attempt failed, and in fact the local Claude install warned [Golan] that the email may be a phishing attack. Where this mildly interesting research takes a really interesting turn, is when he asked Claude if such an attack could ever work.

Claude gave some scenarios where such an attack might succeed, and [Golan] pointed out that each new conversation with Claude is a blank slate. This led to a bizarre exchange where the running instance of Claude would play security researcher, and write emails intended to trick another instance of Claude into doing something it shouldn’t. [Golan] would send the emails to himself, collect the result, and then come back and tell Researcher Claude what happened. It’s quite the bizarre scenario. And it did eventually work. After multiple tries, Claude did write an email that was able to coerce the fresh instance of Claude to manipulate the file system and run calc.exe. This is almost the AI-guided fuzzing that is inevitably going to change security research. It would be interesting to automate the process, so [Golan] didn’t have to do the busywork of shuffling the messages between the two iterations of Claude. I’m confident we’ll cover many more stories in this vein in the future.

Bits and Bytes

SugarCRM fixed a LESS code injection in an unauthenticated endpoint. These releases landed in October of last year, in versions 13.0.4 and 14.0.1. While there isn’t any RCE at play here, this does allow Server-Side Request Forgery, or arbitrary file reads.

Cryptojacking is the technique where a malicious website embeds a crypto miner in the site. And while it was particularly popular in 2017-2019, browser safeguards against blatant cryptojacking put an end to the practice. What c/side researchers discovered is that cryptojacking is still happening, just very quietly.

There’s browser tidbits to cover in both major browsers. In Chrome it’s a sandbox escape paired with a Windows NT read function with a race condition, that makes it work as a write primitive. To actually make use of it, [Vincent Yeo] needed a Chrome sandbox escape.

ZDI has the story of Firefox and a JavaScript Math confusion attack. By manipulating the indexes of arrays and abusing the behavior when integer values wrap-around their max value, malicious code could read and write to memory outside of the allocated array. This was used at Pwn2Own Berlin earlier in the year, and Firefox patched the bug on the very next day. Enjoy!

This Week In Security: Lingering Spectre, Deep Fakes, And CoreAudio

May 16, 2025 by 5 Comments

Spectre lives. We’ve got two separate pieces of research, each finding new processor primitives that allow Spectre-style memory leaks. Before we dive into the details of the new techniques, let’s quickly remind ourselves what Spectre is. Modern CPUs use a variety of clever tricks to execute code faster, and one of the stumbling blocks is memory latency. When a program reaches a branch in execution, the program will proceed in one of two possible directions, and it’s often a value from memory that determines which branch is taken. Rather than wait for the memory to be fetched, modern CPUs will predict which branch execution will take, and speculatively execute the code down that branch. Once the memory is fetched and the branch is properly evaluated, the speculatively executed code is rewound if the guess was wrong, or made authoritative if the guess was correct. Spectre is the realization that incorrect branch prediction can change the contents of the CPU cache, and those changes can be detected through cache timing measurements. The end result is that arbitrary system memory can be leaked from a low privileged or even sandboxed user process.

In response to Spectre, OS developers and CPU designers have added domain isolation protections, that prevent branch prediction poisoning in an attack process from affecting the branch prediction in the kernel or another process. Training Solo is the clever idea from VUSec that branch prediction poisoning could just be done from within the kernel space, and avoid any domain switching at all. That can be done through cBPF, the classic Berkeley Packet Filter (BPF) kernel VM. By default, all users on a Linux system can run cBPF code, throwing the doors back open for Spectre shenanigans. There’s also an address collision attack where an unrelated branch can be used to train a target branch. Researchers also discovered a pair of CVEs in Intel’s CPUs, where prediction training was broken in specific cases, allowing for a wild 17 kB/sec memory leak.

Continue reading “This Week In Security: Lingering Spectre, Deep Fakes, And CoreAudio” →

This Week In Security: OpenSSH, JumbledPath, And RANsacked

February 21, 2025 by 8 Comments

OpenSSH has a newly fixed pair of vulnerabilities, and while neither of them are lighting the Internet on fire, these are each fairly important.

The central observation made by the Qualsys Threat Research Unit (TRU) was that OpenSSH contains a code paradigm that could easily contain a logic bug. It’s similar to Apple’s infamous goto fail; SSL vulnerability. The setup is this: An integer, r, is initialized to a negative value, indicating a generic error code. Multiple functions are called, with r often, but not always, set to the return value of each function. On success, that may set r to 0 to indicate no error. And when one of those functions does fail, it often runs a goto: statement that short-circuits the rest of the checks. At the end of this string of checks would be a return r; statement, using the last value of r as the result of the whole function.

Continue reading “This Week In Security: OpenSSH, JumbledPath, And RANsacked” →

This Week In Security: The UK Wants Your ICloud, Libarchive Wasn’t Ready, And AWS

February 14, 2025 by 15 Comments

There’s a constant tension between governments looking for easier ways to catch criminals, companies looking to actually protect their users’ privacy, and individuals who just want their data to be truly private. The UK government has issued an order that threatens to drastically change this landscape, at least when it comes to Apple’s iCloud backups. The order was issued in secret, and instructed Apple to provide a capability for the UK officials to access iCloud backups that use the Advanced Data Protection (ADP) system. ADP is Apple’s relatively new end-to-end encryption scheme that users can opt-into to make their backups more secure. The key feature here is that with ADP turned on, Apple themselves don’t have access to decrypted user data.

If this order wasn’t onerous enough, it seems to explicitly include all ADP-protected data, regardless of the country of origin. This should ring alarm bells. The UK government is attempting to force a US company to add an encryption backdoor to give them access to US customer data. Cryptographer [Matthew Green] has thoughts on this situation. One of the slightly conspiratorial theories he entertains is that portions of the US government are quietly encouraging this new order because the UK has weaker protections against unreasonable search and seizure of data. The implication here is that those elements in the US would use this newfound UK data access capability to sidestep Fourth Amendment protections of citizens’ data. This doesn’t seem like much of a stretch.

[Matthew] does have a couple of suggestions. The first is passing laws that would make it illegal for a US company to add backdoors to their systems, specifically at the request of foreign nations. We’ve seen first-hand how such backdoors can backfire once accessed by less-friendly forces. In an ironic turn of fate, US agencies have even started recommending that users use end-to-end encrypted services to be safe against such backdoors. Technically, if this capability is added, the only recourse will be to disable iCloud backups altogether. Thankfully Apple has pushed back rather forcefully against this order, threatening to simply turn off ADP for UK users, rather than backdoor the rest of the world. Either way, it’s a scary bit of overreach.

Continue reading “This Week In Security: The UK Wants Your ICloud, Libarchive Wasn’t Ready, And AWS” →

This Week In Security: License Plates, TP-Link, And Attacking Devs

December 27, 2024 by 22 Comments

We’re covering two weeks of news today, which is handy, because the week between Christmas and New Years is always a bit slow.

And up first is the inevitable problem with digital license plates. Unless very carefully designed to be bulletproof, they can be jailbroken, and the displayed number can be changed. And the Reviver plates were definitely not bulletproof, exposing a physical programming port on the back of the plate. While it’s not explicitly stated, we’re guessing that’s a JTAG port, given that the issue is considered unpatchable, and the port allows overwriting the firmware. That sort of attack can be hardened against with signed firmware, and using an MCU that enforces it.

This does invite comparisons to the James Bond revolving license plate — and that comparison does put the issue into context. It’s always been possible to swap license plates. If someone really wants to cause mischief, traditional plates can be stolen, or even faked. What a digital plate adds to the equation is the ability to switch plate numbers on the fly, without stopping or turning a screwdriver. Regardless, this seems like it will be an ongoing problem, as so many manufacturers struggle to create secure hardware.

Malicious RDP

There’s a clever attack, that uses Microsoft’s Remote Desktop Protocol (RDP), to give away way too much control over a desktop. That’s accomplished by sending the target a .rdp file that shares local resources like the clipboard, filesystem, and more. What’s new is that it seems this theoretical attack has now shown up in the wild.

The attack campaign has been attributed to APT29, CozyBear, a threat actor believed to be associated with Russia’s Foreign Intelligence Service. This attribution tracks with the victims of choice, like government, research, and Ukrainian targets in particular. To escape detection, the malicious RDP endpoints are set up behind RDP proxies, running on services like AWS. The proxies and endpoints are accessed through TOR and other anonymous proxies. The .rdp files were spread via spear-phishing emails sent through compromised mail servers. The big push, with about 200 targets, was triggered on October 22nd. Researchers at TrendMicro believe this was the end of a targeted campaign. The idea being that at the end of the campaign, it no longer matters if the infrastructure and methods get discovered, so aim for maximum impact.

Free* Mcdonalds?

Here we learn that while McDonald’s USA dosn’t have a bug bounty program, McDonald’s India does — and that’s why researcher [Eaton Zveare] looked there. And found a series of Broken Object Level Authorization (BOLA) bugs. That’s a new term to this column, but a concept we’ve talked about before. BOLA vulnerabilities happen when a service validates a user’s authentication token, but doesn’t properly check that the user is authorized to access the specific resources requested.

In the McDonald’s case, any user of the web app is issued a guest JWT token, and that token is then valid to access any Order ID in the system. That allows some interesting fun, like leaving reviews on other users’ orders, accessing delivery maps, and getting copies of receipts. But things got really interesting when creating an account, and then ordering food. A hidden, incomplete password login page allowed breaking the normal user verification flow, and creating an account. Then after food is added to the cart, the cart can be updated to have a total price of a single rupee, about the value of a penny.

This research earned [Eaton] a $240 Amazon gift card, which seems a little stingy, but the intent behind the gesture is appreciated. The fixes landed just over 2 months after reported, and while [Eaton] notes that this is slower than some companies, it’s significantly faster than some of the less responsive vendors that we’ve seen.

Banning TP-Link

The US Government has recently begun discussing a plan to ban TP-Link device purchases in the United States. The reported reason is that TP-Link devices have shipped with security problems. One notable example is a botnet that Microsoft has been tracking, that primarily consists of TP-Link devices.

This explanation rings rather hollow, particularly given the consistent security failings from multiple vendors that we’ve covered on this very column over the years. Where it begins to make more sense is when considered in light of the Chinese policy that all new vulnerabilities must first be reported to the Chinese government, and only then can fixes be rolled out. It suggests that the US Commerce Department suspects that TP-Link is still following this policy, even though it’s technically now a US company.

I’m no stranger to hacking TP-Link devices. Many years ago I wrote a simple attack to put the HTTPD daemon on TP-Link routers into debug mode, by setting the wifi network name. Because the name was used to build a command run with bash, it was possible to do command injection, build a script in the device’s /tmp space, and then execute that script. Getting to debug mode allowed upgrading to OpenWRT on the device. And that just happens to be my advice for anyone still using TP-Link hardware: install OpenWRT on it.

Developers Beware

We have two separate instances of malware campaigns directly targeting developers. The first is malicious VSCode extensions being uploaded to the marketplace. These fakes are really compelling, too, with lots of installs, reviews, and links back to the real pages. These packages seem to be droppers for malware payloads, and seem to be targeting cryptocurrency users.

If malware in your VSCode extensions isn’t bad enough, OtterCookie is a campaign believed to come from North Korea, spreading via fake job interviews. The interview asks a candidate to run a Node.js project, or install an npm package as part of prep. Those are malicious packages, and data stealers are deployed upon launch. Stay frosty, even on the job hunt.

Bits and Bytes

PHP has evolved over the years, but there are still a few quirks that might trip you up. One of the dangerous ones is tied up in $_SERVER['argv'], a quick way to test if PHP is being run from the command line, or on a server. Except, that relies on register_argc_argv set to off, otherwise query strings are enough to fool a naive application into thinking it’s running on the command line. And that’s exactly the footgun that caught Craft CMS with CVE-2024-56145.

Australia may know something we don’t, setting 2030 as the target for retiring cryptography primitives that aren’t quantum resistant. That’s RSA, Elliptic-curve, and even SHA-256. It’s a bit impractical to think that those algorithms will be completely phased out by then, but it’s an interesting development to watch.

Fuzzing is a deep subject, and the discovery of 29 new vulnerabilities found in GStreamer is evidence that there’s still plenty to discover. This wasn’t coverage-guided fuzzing, where the fuzzer mutates the fuzzing input to maximize. Instead, this work uses a custom corpus generator, where the generator is aware of how valid MP4 files are structured.

This Week In Security: National Backdoors, Web3 Backdoors, And Nearest Neighbor WiFi

December 6, 2024 by 16 Comments

Maybe those backdoors weren’t such a great idea. Several US Telecom networks have been compromised by a foreign actor, likely China’s Salt Typhoon, and it looks like one of the vectors of compromise is the Communications Assistance for Law Enforcement Act (CALEA) systems that allow for automatic wiretapping at government request.

[Jeff Greene], a government official with the Cybersecurity and Infrastructure Security Agency (CISA), has advised that end-user encryption is the way to maintain safe communications. This moment should forever be the touchstone we call upon when discussing ideas like mandated encryption backdoors and even the entire idea of automated wiretapping systems like CALEA. He went on to make a rather startling statement:

I think it would be impossible for us to predict a time frame on when we’ll have full eviction

There are obviously lots of unanswered questions, but with statements like this from CISA, this seems to be an extremely serious compromise. CALEA has been extended to Internet data, and earlier reports suggest that attackers have access to Internet traffic as a result. This leaves the US telecom infrastructure in a precarious position where any given telephone call, text message, or data packet may be intercepted by an overseas attacker. And the FCC isn’t exactly inspiring us with confidence as to its “decisive steps” to fix things. Continue reading “This Week In Security: National Backdoors, Web3 Backdoors, And Nearest Neighbor WiFi” →