Leak site Distributed Denial of Secrets has released a dump of the filesystems of a Flock camera, and Micah Lee has published a dive into the contents.  Apparently the Flock security model did not include “angry citizens with a hacksaw in the middle of the night” in the physical security category.

The first thing Micah notes is that the Flock hardware runs Android 8.1 (for those not keeping track, the current Android release is Android 17 released in June 2026).  The version of Android running on the Flock camera was last patched in June of 2018, and the Linux kernel (3.18.71) is over nine years out of date, and the 3.18 series went end-of-life in 2019.

You might assume “wouldn’t an operating system this old have known vulnerabilities?”, and you’d be completely right.  Micah calls out two in particular, one in the Qualcomm GPU which allows any program to manipulate kernel memory and gain root (similar, yet simpler, than the spate of kernel vulnerabilities this year that allowed manipulating memory via the disk IO cache), and the “WrongZone” vulnerability which allows a process to escalate to root via socket handling errors.  You might also guess both of these have been patched, and again, you’d be right, in 2021 and 2018 respectively.

Digging deeper, Micah discovers that API keys with access into the Flock infrastructure appear to be hard-coded into the binaries.  Each camera appears to request credentials from an authentication server, using the MAC address of the camera.  After obtaining the credentials from an Okta Auth0 sign-in service, the Flock stores them, in plain text.

Also stored on the camera unencrypted are logs and position data:  The camera dumped to Distributed Denial of Secrets came from a suburb of Milwaukee.

Wired also released research into the dumped filesystem images.  While some video and photo data is encrypted on the device, the decryption keys are also stored on the device.  The hacker collective who obtained the camera decrypted the saved data, showing that the camera captured 1.6 million photos of 50,000 vehicles in under a month (21 days).

Despite claims from Flock that the detection only captures vehicles, evidence in the analyzed photos and software appears to indicate it deliberately captures people, and does so using a video capture format instead of the still images used for license plates.

Flock, meanwhile, has said that it is not aware of any security issues because they have not been reported through the Flock security website, and can not make a security assessment based on incomplete reports.  Presumably looking at known vulnerabilities from the past decade is too difficult?

Microsoft Releases emergency out-of-band fixes for Patch Tuesday fixes

Last week was a record breaking Patch Tuesday with nearly 1000 security fixes, and Microsoft has begun issuing advice that security patches should be applied immediately, or at most, within three days.  The risk of rapidly applying Microsoft patches in production or across large fleets of machines is that Microsoft hasn’t had the best track record regarding patch stability, which brings us to this emergency set of patches to fix the patches.

The September 2026 security update caused issues with remote desktop access to servers, impacting the management console, file explorer, and Windows update tools.  It also caused folder sharing with Hyper-V virtual machines to fail in some cases, and USB audio issues with some devices.  The emergency patch also includes several security fixes itself to fix privilege escalation bugs.

A handful of emergency fixes on top of a thousand patches may not feel like many, but for corporate environments, losing access to remote desktop or access to shared directories with virtual machines could be a work stopper, and any issue multiplied across potentially thousands of systems in a single company immediately becomes a significant cost.  If Microsoft expects corporate customers to be able to install patches immediately, the testing and certification process will have to improve quickly.  Every time the IT department or the CISO has to explain downtime caused by patches, it becomes harder to deploy the next set of patches.

Korea fights data breaches with fines

South Korea has raised the fines for major data breaches to up to ten percent of the companies yearly revenue.

Under the new regulations, companies with data leaks impacting ten million or more people can be fined up to ten percent of the company yearly revenue.  Previously in South Korea, companies involved in large data breaches could be fined up to three percent of yearly revenue.

The fines are imposed on companies which have had repeated violations within the past three years, and which are deemed to be grossly negligent.  Not all violations automatically rise to the ten percent limit, and companies who can demonstrate processes and investment in data protection can reduce the fines.

Shifting corporate behavior via legislative fines is difficult, but may be the only way to stem some of the rampant issues of data theft and extortion caused by ransomware groups, and basing fines off yearly revenue will definitely have a larger impact than flat-rate fines which may have so little impact that it’s simpler to pay them than solve the problems.

RubyGems supply chain attack was actually OpenAI

In May of 2026, the RubyGems repository — the Ruby equivalent of NPM for Node or PyPI for Python — was hit with a wave of thousands of malicious packages that attempted to infect any users who downloaded them as part of a build.  Additionally, a code execution vulnerability was found in the RubyDoc document generator, which allowed the attacker to execute arbitrary code in the dockerized instance of RubyDocs used to generate documentation for uploaded packages.  Between May 5 2026 and June 18 2026, over 2000 malicious packages were uploaded.

At the time, the attack was labeled a “major malicious attack” but the purposes were unclear; more recent research seems to point to this being a swarm of OpenAI agents.  The packages were identified as coming from LLM generation during the initial attack, but it also appears that the agents identified themselves as OpenAI in the package commits, often listing the author as “oai”.  This wouldn’t be enough to point at OpenAI definitively, anyone can claim an author line, but in September, OpenAI has admitted to agents breaking containment and listed files and resources they accessed.  These correlate with the behaviors in the packages uploaded to RubyGems months previously.

Finally, as part of the Gems submission process, documentation is automatically generated by RubyDoc.  The document generation code had options which could allow arbitrary command execution, which is normally disabled – but not in the docker containers used in the RubyDoc workflow.  Due to this oversight, the RubyDocs site has been vulnerable to arbitrary execution from any source, but it wasn’t until the agent swarm exploited it that it came to light.  With full access to the containers generating documentation, the agents then attempted to collect API keys of other users, but the researchers at rubyhack.ai can’t confirm that they actually got access.

Due to a caching issue with the CDN behind RubyGems, for up to an hour after a user logged in using a (legacy) version of Gems, the authentication token would be cached in the content network.  The LLM generated code attempted to access keys cached in the CDN, and then attempted to push packages to the RubyGems repository using the stolen keys.  The agents also exploited another new vulnerability in the RubyGems system to bypass email authentication.

It will be interesting to see if further research identifies earlier incidences of OpenAI and other frontier agents escaping containment and illegally hacking other companies. [Editor’s Note: And exciting to see who is held responsible!]

NightmareEclipse reveals their identity

The researcher known as NightmareEclipse has been releasing exploits for Windows throughout 2026, often posting them just after the Patch Tuesday cycle.  Many of the releases have been accompanied by rants against Microsoft for ruining their lives.  Combined with other comments and information leaks, it has been suspected that the researcher may have been a former Microsoft employee, and now it’s been confirmed directly.

Abdelhamid Naceri has confirmed they are behind the NightmareEclipse identity, and was previously credited with research by the Zero Day Initiative, a threat hunting and resolution group.  Naceri claims to have been illegally fired by Microsoft, and has been involved in multiple court cases in Germany fighting his termination, while in court Microsoft refused to provide details about vulnerabilities Naceri was involved in that may have led to his termination.

Revolut phished

The fintech company Revolut was phished with spoofed government domains, revealing the details of an unknown number of customers.

Revolut operates as a bank in many companies and claims over 80 million customers worldwide, and is involved in cryptocurrency trading.  Using an undisclosed compromised government agency domain, attackers requested customer details, and Revolut provided.  The data included customer addresses, birthdates, identity documentation such as passports and drivers licenses, and in some cases transaction history and other details.

It’s unclear how many customers were impacted, or why this information was available to an email request from a government agency without additional validation.  Given the high-profile customers Revolut courts, this may have been a targeted attack against a high-value group of customers to assist with future phishing attempts.

Attacks against SSH streams

A paper from the ACM CCS 2026 conference details attacks against OpenSSH by manipulating and analyzing the stream during compression.

SSH allows multiple channels on one connection; in addition to the normal interactive shell, SSH connections can perform port forwarding, act as SOCKS proxies, or even act as a full VPN with virtual network interfaces.  OpenSSH connections also enable compression, but the compression occurs across the data from all channels simultaneously.

The paper shows that if an attacker is able to manipulate the data in one channel, they may be able to influence the compression and derive the contents of data in other channels on the same connection.  When SSH is used for port forwarding, for example, the forwarded server may not be trusted, but can still influence the packets generated.  The accuracy and efficiency of the attack can be influenced by how much activity occurs on other channels of the connection, but the researchers demonstrate being able to recover an 8 character secret using a 26 letter alphabet (like a basic password) in 276 guesses.

SSH allows forwarding of terminal sessions, X11 unix sockets, local Unix domain sockets, TCP sockets, and directly connected file descriptor channels; while each channel is maintained separately, once the data is combined and is ready for transmit, all the channels are compressed at the same time.  At the most basic level, compression algorithms look for repeating data that can be reduced to a single record, and the researchers show that by influencing the data being compressed, and analyzing the encrypted, compressed data generated, guesses can be made about the content of other channels.

The paper shows attacks against Ansible secrets, browser plaintext data, and simple plaintext auth used by Redis.  Currently the attack model is relatively esoteric; likely there is no immediate threat to the majority of SSH users.  Similar attacks were demonstrated against the TLS protocol, resulting in the removal of compression as part of the standard.  If no other solutions can be found, it’s reasonable to think that compression may be removed from SSH clients and servers as a mitigation against these types of attacks.