Agregátor RSS

Seems like as good a time as any to upgrade older hardware

There are early indications of active attacks targeting end-of-life Zyxel NAS boxes just a few weeks after details of three critical vulnerabilities were made public.…

Cybersecurity researchers have detailed a now-patched security flaw affecting the Ollama open-source artificial intelligence (AI) infrastructure platform that could be exploited to achieve remote code execution. Tracked as CVE-2024-37032, the vulnerability has been codenamed Probllama by cloud security firm Wiz. Following responsible disclosure on May 5, 2024, the issue was addressed in version

Cybersecurity researchers have detailed a now-patched security flaw affecting the Ollama open-source artificial intelligence (AI) infrastructure platform that could be exploited to achieve remote code execution. Tracked as CVE-2024-37032, the vulnerability has been codenamed Probllama by cloud security firm Wiz. Following responsible disclosure on May 5, 2024, the issue was addressed in version Newsroomhttp://www.blogger.com/profile/[email protected]

Apple’s decision not to introduce Apple intelligence, iPhone Mirroring, or SharePlay Screen Sharing in the European Union this year isn’t surprising, and reflects concerns around privacy far more than being a response to Europe’s decision to act against Apple’s App Store compliance.

The news basically is that Apple has confirmed it will not introduce the Apple Intelligence features it announced at WWDC in EU nations because it has concerns around the application of the Digital Markets Act (DMA). 

Apple Intelligence delayed, App Store gets a DMA slap

“Due to the regulatory uncertainties brought about by the Digital Markets Act, we do not believe that we will be able to roll out three of these [new] features — iPhone Mirroring, SharePlay Screen Sharing enhancements, and Apple Intelligence — to our EU users this year,” Apple told the Financial Times.

At the same time Apple made its announcement, the EU itself announced it will begin to take enforcement action against the company for breach of the DMA. Europe is concerned about elements of Apple’s offer to developers and the fees it charges the small number of developers who are the most successful on the store, arguing this stifles competition. Apple says it has made changes and is “confident” its plans align with that law.

It’s worth noting that the EU recently proposed an incredibly intrusive surveillance law that would break end-to-end encryption. While it looks like those proposals may have been shelved, Apple might have decided to stall while it waits to see what kind of shabby surveillance laws do get passed. 

Privacy or convenience? It’s up to EU

If you think about it, the beauty of Apple Intelligence is that it uses information your device has collected about you in order to function. But the risk of that information existing — even on your device — is that under the DMA, it’s not certain the EU won’t insist on that data, your data, being opened up to competitors. 

That’s a lot of information.

Apple is committed to keeping that information private and secure, but once it exists and is on the device in some form, I expect the company is concerned the DMA could force it to open up the information to third parties who want to compete with their own AI. As we’ve seen since the invention of the Internet, not every company is legitimate, ethical, or trustworthy, and even those that are might not have enough clout to invest in the world’s best security teams to maintain safety on their platforms.

A Pandora’s box chock-full of trouble

I get the sense that Apple’s decision to hold back on Apple Intelligence in the EU reflects the ongoing battle between the two entities as Europe forces Apple to open up a little. Given the source of this speculation, that might be correct, but the analysis misses what’s really at stake: once you have all your personal information turned into usable data on your device, every spook, hacker, fraudster, blackmailer, censor, despot, cop, or secret service operative is going to want to take a look.

That means any weakness in protecting that information opens a Pandora’s box of misanthropy — affecting consumers, corroding trust, and enabling surveillance at a scale no one has seen ever before on our sadly ailing planet. 

Could the EU end up without Apple Intelligence? 

It feels possible EU might never get Apple Intelligence.

Apple says: “We are committed to collaborating with the European Commission (EC) in an attempt to find a solution that would enable us to deliver these features to our EU customers without compromising their safety.”

According to the Financial Times, Apple seeks “clarity” from the European Commission regarding the level of access it would need to grant to third parties over Apple Intelligence features in order to be DMA-compliant.

However, rather than providing any insight into requirements, an EU rep said companies like Apple are, “Welcome to offer their services in Europe, provided that they comply with our rules aimed at ensuring fair competition.” Which is, of course, what Apple is asking for, it just wants to know how those rules will be applied to its service before launch, rather than working with decisions made after the event.

What isn’t yet clear is the extent to which other AI providers might be affected. Is it possible the European Commission might have just created an obstacle to AI deployment?

Who has the energy for this?

And, of course, the big conversation everyone should be having concerning artificial intelligence is one Europe’s regulators don’t appear to be addressing at all — the energy consumption of AI servers. Combined, the world’s data centers now consume more power in a year than the entire Italian nation, and this is set to increase exponentially. Perhaps waiting until privacy, security, and energy challenges are solved makes sense after all?

One more thing is also certain: That with the removal of these three features the temptation to upgrade to iOS 18 among European users will be lower than ever before, given they comprise the majority of improvements to the OS.

Please follow me on Mastodon, or join me in the AppleHolic’s bar & grill and Apple Discussions groups on MeWe.

Předseda výboru Sněmovny reprezentantů pro zpravodajské služby Mike Turner obvinil Bidenovu administrativu ze zatajování podrobností o stavu ruského protisatelitního zbraňového programu. Aktuální situaci označil za novodobý ekvivalent kubánské raketové krize. Byl to právě Turner, kdo v únoru jako ...

Morpheus comms system online by 2025? You must be dreaming

The UK government has been accused of blowing £174 million ($220 million) on "external advice" for a new radio system for the armed forces that has been beset by delays and cancelled contracts.…

Počet útoků ransomwaru na cíle ve zdravotnictví každoročně roste. To má však zásadní důsledky pro zdraví a pohodu pacientů v postižených zařízeních. Závažné problémy tohoto trendu naplno odhalily nedávné útoky na londýnské nemocnice. O problému informují deníky New York Times a Wired. Británie ...

Začátky jsou vždy těžké: první notebook s RISC-V spatřil světlo světa v závěru roku 2022 a dlouho byl na trhu sám. Teď má nástupce a dokonce i nějakou konkurenci.

Podle zdrojů tradičně dobře informovaného Marka Gurmana z Bloombergu v Applu aktuálně připravují dvoje další chytré brýle. Známe jen pod kódovými označeními N109 a N107. Ty první budou Vision Pro druhé generace, které mají přinést vyšší výkon, lepší kamery nebo lehčí a pohodlnější tělo. Původně se ...

Learn about critical threats that can impact your organization and the bad actors behind them from Cybersixgill’s threat experts. Each story shines a light on underground activities, the threat actors involved, and why you should care, along with what you can do to mitigate risk.  Cybersecurity professionals are facing unprecedented challenges as they strive to manage increasing workloads

Learn about critical threats that can impact your organization and the bad actors behind them from Cybersixgill’s threat experts. Each story shines a light on underground activities, the threat actors involved, and why you should care, along with what you can do to mitigate risk.  Cybersecurity professionals are facing unprecedented challenges as they strive to manage increasing workloads The Hacker Newshttp://www.blogger.com/profile/[email protected]

Minimální telefon by měl stát i minimum peněz • Light Phone III slibuje osvobození od barev a sociálních sítí • S cenou, která se rovná iPhonu, je to ale zbytečně drahý špás

V neděli 30. června skončí (EOL) podpora CentOS Linux 7.

Ve více než stovce poboček McDonald's v USA testovali nasazení umělé inteligence na zdánlivě jednoduchý úkol: přijímání objednávek od zákazníků drive-thru. AI chatboti měli klientům i obsluze usnadnit a zrychlit objednávkový proces. Ačkoli výsledky nebyly špatné, řetězec rychlého občerstvení se ...

A busy few days for security teams

There were data breaches galore in the US last week with various major incidents reported to state attorneys general, some in good time, some not.…

Part 1: XZ backdoor story – Initial analysis
Part 2: Assessing the Y, and How, of the XZ Utils incident (social engineering)
Part 3: XZ backdoor. Hook analysis

In our first article on the XZ backdoor, we analyzed its code from initial infection to the function hooking it performs. As we mentioned then, its initial goal was to successfully hook one of the functions related to RSA key manipulation. In this article, we will focus on the backdoor’s behavior inside OpenSSH, specifically OpenSSH portable version 9.7p1 – the most recent version at this time.

To better understand what’s going on, we recommend you to read Baeldung’s article about SSH authentication methods and JFrog’s article about privilege separation in SSH.

Key findings

Our analysis revealed the following interesting details about the backdoor’s functionality:

• The attacker set an anti-replay feature to avoid possible capture or hijacking of the backdoor communication.
• The backdoor author used a custom steganography technique in the x86 code to hide the public key, a very clever technique to hide the public key.
• The backdoor hides its logs of unauthorized connections to the SSH server by hooking the logging function.
• The backdoor hooks the password authentication function to allow the attacker to use any username/password to log into the infected server without any further checks. It also does the same for public key authentication.
• It has remote code execution capabilities that allow the attacker to execute any system command on the infected server.

Detailed analysis

There are three functions that the backdoor attempts to hook, of which RSA_public_decrypt is the primary target and RSA_get0_key is the secondary. The third function, EVP_PKEY_set1_RSA, doesn’t exist in the SSH server version in question. It may be an artifact left over from the tool used for malicious public key generation (this function is used by an independent ssh-keygen tool included in the OpenSSH packet), or it may have been used in a rare or outdated version of the SSH server.

The two target functions in the latest SSH server version are called when the RSA certificate is configured as an SSH authentication method. They first check if an incoming RSA connection uses authentication data (RSA key) as an argument. If so, the backdoor passes it to a common function (called by all hooks) that parses this RSA key and extracts information that is embedded in its modulus part. The backdoor’s main payload function works only once during a client preauth session, when the RSA-based authentication checks are performed.

RSA_public_decrypt hook function

An attacker must generate a specific RSA key to interact with the backdoored server; the key is used as a container for the attacker’s commands in SSH connections using CA certificates.

The RSA key is represented by a structure in the OpenSSL library that contains the E (exponent) and N (modulus). The backdoor extracts and processes the RSA modulus, which means that the malicious payload is packed inside the N value from the RSA cryptosystem.

The custom RSA modulus must conform to the following format to be processed correctly by the backdoor:

RSA modulus data structure

There are three fields in the payload header (PartialCommand1, 2 and 3 in the scheme above) that are used to calculate the command type and also act as a form of magic number check. The command type is calculated using the following formula: PartialCommand3 + (PartialCommand2 * PartialCommand1), where the result of the calculation must be a value between 0 and 3:

Command type calculation

If the calculated check passes, the code proceeds to the payload decryption and payload signature check.

ED448-encrypted public key extraction – x86-based steganography

To decrypt and verify the payload data, the backdoor uses an ED448 public key extracted from the binary.

When we first encountered the key extraction procedure, it looked like the backdoor authors had managed to create code that generated a correct public key before the private key, which should be impossible. Normally, for the Elliptic Curve Algorithm, the private key must be generated first, and then the public key is calculated from it. To solve the mystery of generating the public key from the binary, we analyzed the source code of various cryptographic libraries and came up with nothing. We then analyzed the backdoor code more closely, and found that the keys were generated using a regular procedure. However, the attackers used a custom steganography technique in the x86 code to hide an arbitrary message (in this case, the public key).

The public key information was scattered inside the binary code within specific valid instructions. The method of recovering the key is somewhat similar to the gadget scanning technique in a return-oriented programming (ROP) binary exploitation scenario. But here the “gadgets” are actually register-register instructions (e.g., mov rdi, rbx), each of which holds one bit of information, whose value is either 1 or 0.

To achieve key recovery, some functions, usually at the beginning of the function, call the “key rebuild” algorithm with specific arguments.

Partial key rebuild function call

The arguments used by this algorithm are:

• BitIndex: the starting value that holds the current key index to be decoded and also holds which bit should be set initially in the encrypted key bitmap.
• Total Instructions: the number of register-register instructions to scan in the current function.
• Key Index: the specific key index this function will work to reconstruct. This value exists to avoid rescanning the same function if it is called a second time.

Register-register instruction decoding

The key rebuild algorithm scans certain functions of the backdoor from beginning to end looking for register-register instructions. When it finds an instruction, it decodes the ‘BitIndex’ value to extract the correct byte index and bit to be set.

Encrypted key rebuild code snippet

The BitIndex value is unpacked to determine the target index in the buffer. It then adds (bitwise or) the bit to the current value at that index. As the encrypted public key buffer is initialized with zeros, the rebuilder algorithm will only activate specific bits inside it. It sets the key bit value to 1 if the register-register instruction matches the opcode criteria (image above), or skips it, indicating that this bit value should remain zero. After that, the BitIndex value increases.

The algorithm determines whether the bit should be set or not for each instruction individually, even if the instructions have the same disassembly representation. This is because some instructions can have the same assembly code but different opcodes.

Public key rebuild algorithm

In general, for each instruction found, the BitIndex is used to reconstruct a specific part of the encrypted key. In total, 456 instructions are hunted through the binary execution, and the encrypted public key is rebuilt by the end of this process.

Key rebuild automation

In our research, we recreated the entire key rebuilding process that results in the encrypted public key that is later decrypted.

Payload decryption and signature check

The ED448 public key is encrypted using the ChaCha20 algorithm, where the key and nonce are the result of ChaCha20 encryption of a buffer consisting of zeros, with zeros used as the key and nonce.

After decryption, the backdoor takes the first 32 bytes of the public key and uses them as the key to decrypt the payload body, which is also ChaCha20 encrypted.

Backdoor payload decryption and check diagram

Payload signature check

The decrypted payload contains the signature of the remaining data in its header. To verify the signature, one must have a private key to sign the payload. In the expected attack scenario, only the backdoor author would have access to sign and send payloads to the infected server.

To verify the integrity and authenticity of the payload, the backdoor again uses the decrypted ED448 public key to confirm that the incoming payload was signed with the attacker’s private key.

Payload integrity and authenticity checks

It also takes the SHA-256 hash of the server’s public key (taken from the initial SSH connection when the server sends the public key) into the payload signed data and verifies that it matches the currently running server. This is done to prevent replay attacks, where a researcher could capture the backdoor communication and replay the same backdoor command to another server.

Anti-replay attack diagram

If all the checks pass, the code proceeds to parse the arguments of the desired backdoor command. The backdoor can execute the commands in two modes, root and non-root, and the execution can vary depending on the privilege level. However, the non-root mode operations don’t appear to be the attacker’s goal, so we’ll describe what the root-mode code does.

Backdoor commands

The command chosen by the attacker depends on the result of the calculation on the header fields. The core backdoor commands essentially allow the attacker to log into the server as root or a regular user and execute some system commands. This section describes what each command does.

Bypass SSH authentication

Both commands 0 and 1 enable root login on the SSH server if it wasn’t previously enabled. Additionally, they can optionally disable the use of Pluggable Authentication Modules (PAM). Next, depending on the command, one of two OpenSSH functions can be hooked, the mm_answer_authpassword if the command is 1 or the mm_answer_keyallowed if the command is 0.

First command core code

The mm_answer_authpassword hook allows the attacker to use any username/password to log in to the infected server without any further checks. We’ll describe mm_answer_keyallowed in more detail later.

Remote command execution via ‘system’ call

If the command is 2, it extracts the command string from the payload data and executes it on the machine using the system function. The attacker can send specific flags in the payload body to specify the user and group ID under which the command should be executed.

Remote command execution code snippet

The last backdoor command (3) simply closes the pre-auth session by exiting the Monitor process.

After the main hook function completes, the backdoor replaces the incoming RSA public key with an empty one to avoid logging any malicious data from the attacker.

The mm_answer_keyallowed hook

If the backdoor command is 0 the mm_answer_keyallowed function will be hooked. This function is called by the privileged Monitor process when the unprivileged NetworkProcessing sshd process receives the publickey method of the ssh-userauth request, allowing the backdoor to process additional commands at that stage.

In this processing method, similar steps occur: the modulus of the attached to the ssh-userauth request RSA key is extracted, decrypted, and its signature is checked as explained earlier. The possible subcommands are:

• SubCommand 1:
  • Hook the mm_answer_authpassword function to allow user login requests with any password.
• SubCommand 2:
  • Hook the mm_answer_keyverify function to allow authentication with fake keys and store the received backdoor command inside an internal structure. The mm_answer_keyverify hook essentially allows the incoming key to pass through the authentication process without further checks, successfully completing authorization with a fake key.
• SubCommand 3:
  • Similar to the system execution method, this command allows the execution of commands using the system function, with the option to choose the group/user ID.

Log hiding capabilities

The backdoor can also hide logs that indicate unauthorized connections to the SSH server by hooking the logging function and optionally calling the libc setlogmask(0x80000000), which causes all Syslog messages to be ignored. 0x80000000 in the argument actually means setting the log mask to zero, because the zero-mask argument is a reserved value with special meaning.

The hook will also filter out log messages by searching them in the prefix tree, as described in our first post about XZ backdoor. Its further behavior depends on which of the targeted messages were found (if any):

Log filtering code snippet

The available log filters are:

Log message How it is processed “Connection closed by “ Temporarily restores libc’s syslog mask to its default value 255, allowing all syslog messages if it was previously cleared, and allows this message to be logged.
Disables syslog messages again by clearing the log mask “Accepted password for ”
“Accepted publickey for “ Replaces these successful connection messages with messages about failed authentication attempts. Also temporarily enables and then disables the syslog mask if it was previously cleared. All other log messages Filtered out (not printed) Conclusion

After three posts on this backdoor, we can conclude that it is indeed a highly sophisticated threat with many peculiarities. Several highlights make this threat unique, such as the way the public key information is embedded in the binary code itself, complicating the recovery process, and the meticulous preparation of the operation, which involves a long-running social engineering campaign.

It is notable that the group or attacker behind this threat has extensive knowledge of the internals of open-source projects such as SSH and libc, as well as expertise in code/script obfuscation used to start the infection.

Kaspersky products detect malicious objects associated with the attack as HEUR:Trojan.Script.XZ and Trojan.Shell.XZ. In addition, Kaspersky Endpoint Security for Linux detects malicious code in sshd process memory as MEM:Trojan.Linux.XZ (as part of the Critical Areas Scan task).

Data centers are draining more electricity from global power grids than ever before because of generative artificial intelligence (genAI) and general AI processing needs. 

The compute capacity to train large language models, the platforms on which generative AI (gen) and AI run, is now roughly doubling every nine months, according to Epoch AI, an AI research institute. The International Energy Agency forecast that global data center electricity demand will more than double from 2022 to 2026, in large part because of AI and cryptocurrency.

That insatiable demand for energy has tech companies scrambling for alternative sources of energy as well as ways to reduce the energy needs of AI technologies.

One potential emerging solution to the AI-compute dilemma is quantum computing, which vastly surpasses today’s binary computing systems in processing capabilities and energy consumption. Studies have shown quantum computing can increase the performance of AI neural networks for tasks such as natural language processing and image analysis.

“Quantum computing definitely augments the power of AI. For example, AI and quantum computing used together can accelerate drug discovery and personalized pharmaceuticals by years. Quantum computing supports AI-based simulation of clinical drug trials so that the trials take one hour instead of ten years,” said Avivah Litan, a vice president analyst at Gartner.

For example, in February, Insilico Medicine, Zapata AI, and the University of Toronto announced they’d demonstrated the first instance of a generative model running on quantum hardware outperforming state-of-the-art classical models in generating viable cancer drug candidates.

What is quantum computing?

In classical computers, bits programmed as units of data have a possible value of one or zero — hence the term binary code. In quantum computers, data units are programmed with quantum bits, known as qubits, which can represent a one, a zero, or a combination of both zero and one at the same time. At a high level, that trait enables quantum computers to be faster and better at fundamental processing tasks than data processing on classical computing systems that use GPUs or CPUs.

For example, Google’s Quantum AI division built a supercomputer based on its Sycamore quantum processor. Each chip currently holds 70 qubits and can reportedly complete in seconds what would take a CPU- or GPU-based supercomputer of similar size decades to process.

From left to right, Google’s rendition of its Quantum computing platform and its Sycamore quantum processor.

Google

“Quantum artificial intelligence with better algorithms… are faster and more accurate,” CompTIA, a global, nonprofit IT association, stated in a blog.

Commercial quantum platforms, such as Microsoft Azure Quantum, AWS Braket, Google Cirq, and others, allow cloud providers to use quantum comuting as compute service offerings.

“Think of these platforms as quantum computing marketplaces whereby the cloud service providers have partnered with multiple quantum computing vendors to provide access to their hardware, software, QSDKs [Quantum software development kits], etc.,” said Heather West, a research manager with IDC.

“Most of these cloud service providers have not, and thus do not, provide access to their own quantum systems, the exception being Google. AI is not a part or related to these offerings,” she added.

As with any technology, along with the positives there are negatives associated with quantum computing. For example, quantum computing poses a serious threat to the cybersecurity systems relied on by virtually every company, according to CompTIA. The current standard for encryption algorithms, such as RSA or SSL/TLS, relies on the complexity in factoring large numbers into primes, and that’s the type of problem quantum computers are great at solving, CompTIA said.

Startups and established companies continue to accelerate their advances in the quantum computing space. Big tech companies such as Alibaba, Amazon, IBM, Google, and Microsoft have already launched commercial quantum-computing cloud services. Two years ago, Goldman Sachs said it planned to introduce quantum algorithms to price financial instruments as soon as 2026. Honeywell anticipates that quantum will form a $1 trillion industry in the decades ahead.

Quantum computing, meet genAI

Some say quantum computing is a natural partner for genAI and can reduce its energy demands.

For example, Sumitomo Mitsui Trust Bank in Japan is using quantum computing to run genAI-powered programs for financial simulation models of future market movements. The bank partnered with Zapata AI, a genAI company that was spun out of Harvard University’s quantum computing lab in 2017.

Christopher Savoie, Zapata AI’s CEO, sees linear algebra (quantum math) as the solution to perform all kinds of AI tasks, including chatbots such as ChatGPT.

“We’re throwing an obscene the amount of GPU energy at chatbots right now. Are we getting that much business value right now from it? We’re hitting a wall: when are we going to make money with that?” said Savoie, who is a molecular biophysicist.

Savoie pointed to Zapata’s research with Insilico Medicine and the University of Toronto to develop cancer drug candidates using a generative model running on quantum hardware.

“When we used this quantum-based model… we were able to develop cancer drugs the other models didn’t,” Savoie said. “We used quantum models to determine what drugs would block this cancer protein and then non-quantum models. The quantum models found two capable drugs that we synthesized and showed they blocked the cancer protein.

“So, it’s qualitatively better,” he continued. “It’s cheaper, faster, and better — better in that we get faster answers. That’s important in drug discovery. You’re saving a lot of money for pharmaceutial companies if you get your answer the first run around. Or you have a more accurate modeling of trading behavior for a bank.”

Zapata AI’s Orquestra platform was specifically designed to run any AI or machine learning model, including more traditional neural networks as well as the company’s proprietary tensor networks.

Tensor networks can be used to model any quantum circuit and run it on today’s classical computers, giving users an on ramp to the potential benefits of future quantum computers, according to Zapata AI. Tensor networks also come with their own advantages for AI today, including more accurate, efficient, and expressive AI models.

“Every quantum circuit can be written as a tensor product, which means we can do things on GPUs that quantum computers will eventually be faster at doing. Zapata and others have shown that quantum math is better at getting better answers in the context of generative AI,” Savoie.

Specifically, Savoie said, quantum statistics can enhance genAI models’ ability to extrapolate missing information and generate new, high-quality information from big data. Generating genuinely new and high-quality data is very important for industrial use cases, he said. 

Early days yet

IDC’s West said quantum computing fits with complex problem solving, but it’s “not a big data solution.” Quantum computing will be useful for solving specific types of problems, she said.

In quantum computing, a qubit begins in a binary state of 0 or 1, but through a process known as annealing, the qubits become entangled, allowing them to represent many possible answers, always with minimum energy. The process occurs in microseconds.

“Quantum annealers are best suited for optimization problems,” West said. “The complex algebraic/factorization problems include some QML [quantum machine learning] problems, but not all AI problems will be suitable for quantum. Research is being conducted to determine how to integrate AI into [quantum computing] and [quantum] into AI to optimize the compute resources needed to solve some of these problems.”

In large part, quantum computing is in very early stages of development, West noted. That’s because the hardware still needs considerable improvements for gate-based models that allow for the​ execution of quantum‌ algorithms. By applying various gates ​sequentially, complex computations can be carried out.

“There are not any real-world applications for this type of system,” West said. “These systems are only useful for small-scale experimentation and debugging. Quantum [computing is] currently being used for solve some scientific and business optimization problems. It is still too early for the integration of AI. Right now, it is only a hypothetical and experimental.”