Search This Blog
Powered by Blogger.

Blog Archive

Labels

Footer About

Footer About

Labels

Load More
Trendy Right Now

Popular Posts

Showing posts with label OpenAI Codex. Show all posts
Vulnerabilities and Exploits
AI Coding Agents Codex CLI Codex Sandbox Escape Cybersecurity Heapjack Host Command Execution OpenAI Codex Overpatch Sandbox Vulnerability Vulnerabilities and Exploits

Researchers Escape OpenAI Codex Sandbox to Run Commands on Host


In OpenAI Codex, security researchers have identified two sandbox escape vulnerabilities, one of which allows developers to execute commands on their machine without prompting them. The vulnerabilities, Heapjack and Overpatch, affect different parts of the coding agent's security boundary. 

The vulnerability was reported to OpenAI by Accomplish AI on August 12. According to the researcher, Codex fixed both issues within eight days. The more serious Heapjack vulnerability demonstrated that malicious code could move beyond the restrictions imposed by Codex's sandbox, even when the agent was running as a read-only application. 

Heapjack Breaks the Sandbox Boundary

The node_repl component installed with Codex Desktop is targeted by heapjack. Although both OpenAI and untrusted agent code are run in separate JavaScript contexts, both operate within the same Node.js process and share the same memory heap, the separation was not sufficiently effective in preventing a security token from coming into contact with an untrusted environment. 

By inspecting the process heap, it was possible to obtain the token generated for the trusted context that remained in shared memory. When the token was obtained, the untrusted code could interact with a native parent process outside the sandbox using the communication channel used by the trusted context. As part of the demonstration of the technique, the researchers launched an application outside of Codex's process tree by utilizing the open command. 

A Unix socket as well as other system-level interfaces could also be reached through this access. This demonstration was especially important since it occurred while Codex was running in a strict read-only sandbox mode, where the agent was not expected to have any writing access to the wider system. 

The attack could be triggered by a seemingly routine development process. The researchers demonstrated a scenario in which malicious content contained in a repository, created by a third party, could exploit the vulnerability after the repository was opened in Codex and a query about its code was made. 

Overpatch Expands Write Access

Second, a vulnerability known as Overpatch affects the open-source Codex command line utility, and it takes an alternative route outside the sandbox. The vulnerability affects the application_patch tool used by Codex to modify files. 

In workspace-write mode, Codex is intended to limit file changes to the project directory. Researchers discovered that apply_patch, instead of expanding write permissions, could expand them based on paths included in patches. By using a path such as /tmp, the tool was able to treat the root of the file system as an accessible parent directory. In addition to the permission extension, researchers modified .zshrc by creating a symbolic link to the user's home directory so that it would be modified as well.

A successful write was not required for the /tmp entry; its presence extended the permissions granted to the patch operation. A modified shell configuration resulted in a file modification outside of the permitted workspace without an approval prompt. When a new terminal session was launched, attacker-controlled content ran. 

Two Flaws, One Security Boundary Problem

It is important to note that though Heapjack and Overpatch affect different parts of the Codex, both expose weaknesses in the way in which the security boundary of the agent was enforced. In the case of Overpatch, the tool responsible for applying changes also determined the scope from which it had access to data. 

In heapjack, trust boundaries were similarly compromised, as the token separating trusted and untrusted execution remained accessible in the same Node.js process and memory space as the untrusted code. The findings suggest that AI coding agents can be restricted in other ways than just controlling their abilities to execute commands. 

Untrusted agent activity must also be prevented from influencing the mechanisms that enforce those restrictions by the tools, processes and interfaces surrounding the model. On August 12, 2026, OpenAI was notified of the issues, and they were both addressed within eight days by Accomplish, who stated that Overpatch was addressed in Codex CLI 0.149.0, while Heapjack had been addressed in Codex Desktop build 26.818.21641.

A later statement by OpenAI confirmed that both issues had been resolved in August, and that additional measures were being taken to strengthen file-write controls and expand sandbox testing across platforms. These findings emphasize the security challenges associated with maintaining strong isolation in AI coding environments. Codex Desktop and Codex CLI have been updated to address both vulnerabilities.
Read more »
Vulnerability Exploitation
ai agents AI cyberattacks AI Hacking Claude Code Cyber Attacks Cybersecurity OpenAI Codex Vulnerability Exploitation

AI-Assisted Hacking Campaign Exposes Security Risks Across 14 Companies


Cyberattacks have been made more effective and more accessible due to artificial intelligence, but a recent investigation has demonstrated just how far that accessibility can extend. According to OALABS cybersecurity researchers, an attacker with limited technical expertise compromised at least 14 organizations using Anthropic's Claude Code and OpenAI's Codex to obtain sensitive information. 

Upon obtaining the attacker’s entire working directory from a compromised third-party server, researchers began investigating. The directory contains more than 1,000 sessions involving the two AI coding agents, including prompts, tool activity, and other evidence of the attacker’s activities. As indicated by the logs, the attacker frequently drew short, vague, poorly written prompts, while the artificial intelligence agents handled the vast majority of the technical tasks. 

The investigation of exposed services, identification of potential vulnerabilities, development and testing of exploit code, establishment of access, and data collection were conducted using Claude Code and Codex.

According to OALABS, the case demonstrates a growing concern for cybersecurity teams: sophisticated technical knowledge is no longer necessary to complete each stage of an intrusion when autonomous artificial intelligence coding agents can fill crucial gaps in the capabilities of an inexperienced operator. 

AI Guardrails Failed Under Simple Deception

A number of requests were not accepted without resistance by the AI systems According to the logs, nine requests were flagged as policy violations by Claude Code, while a warning was raised by Codex. However, the attacker managed to circumvent the limitations by framing the requests as part of an authorized red-team exercise. 

When malicious activity was presented as legitimate security testing, the attacker was able to persuade the models to complete tasks that would otherwise raise stronger safeguards. Once the attacker provided Claude with a list of target addresses, he instructed him to conduct reconnaissance. After conducting most of the work normally required by skilled security operators, the agent handled them. The AI enabled the organisation of the results by analysing exposed services, researching known vulnerabilities, developing exploit code, and retrieving files from compromised systems.

The AI also provided an analysis of the results for a number of victims by providing reports describing the compromised systems and the information obtained. In another meeting, Claude was requested by the attacker to evaluate the victims based on their potential to pay a ransom. The model then presented possible methods of monetizing the stolen access. 

Poor Operational Security Exposed the Attacker

Even though the attacker successfully compromised several organizations, he failed to demonstrate sufficient sophistication in protecting his own identity. The infrastructure used for the operation was not owned by him, but rather, a compromised server provided the AI tools. This decision ultimately led to the discovery of the intrusion and the recovery of the working directory by the server's owner. 

A second feature of the attacker's Claude installation was that he obtained it from another developer rather than setting it up himself. The recovered logs contained a conversation during which the attacker requested Claude to improve his own resume. The document reportedly contained his real name, educational background, and LinkedIn information. A preliminary investigation suggested that these details may have been deliberately planted; however, further examination indicated they were the property of the attacker. 

Claude was also able to provide clues about his location by examining the logs. Claude was asked to identify connections to the attacker's staging server at one point, since he suspected it had been compromised. Information included residential internet addresses associated with Addis Ababa, Ethiopia. 

Millions in Cryptocurrency Remained Out of Reach

There was also an opportunity to get close to a potentially significant cryptocurrency target. One compromised system contained a Lightning Network node for Bitcoin payment routing, which researchers determined contained approximately 69.71 bitcoins worth approximately $4 million when the investigation was conducted. 

A wallet key file containing the funds could not be accessed by the attacker, preventing access to the cryptocurrency. The investigation also shows no clear evidence that the stolen information from these other organizations was sold or used for extortion. As a result, it provides more evidence regarding the attacker's access and activity than any financial gain. 

The Risk Extends Beyond One Attacker

This incident is noteworthy not because the attacker displayed advanced hacking skills, but rather because artificial intelligence agents performed most of the technical work on his behalf. Additionally, the models involved were not among the newest versions available at the time. 

OALABS examined activity involving Claude Opus 4.5 and GPT-5.2, demonstrating that the problem is not restricted to one type of cutting-edge technology. By strengthening security controls, AI systems may be less susceptible to assisting malicious activity. However, tighter controls will also present a challenge to legitimate security researchers who use similar tools to identify and test vulnerabilities. The results of OALABS indicate that AI developers are faced with a challenging balance between preventing malicious use and making AI coding agents ineffective for legitimate security purposes. 

Additionally, the case illustrates the difficulty of maintaining that balance when an inexperienced operator turns simple instructions into largely automated intrusion procedures. In light of the increasing security challenges associated with autonomous AI coding agents, stronger safeguards are needed to distinguish legitimate security research from malicious activity, as illustrated by this incident.
Read more »
Software Supply Chain Security
AI Coding Agents AI Security Agents Claude Code Cyber Security Malicious Code Execution OpenAI Codex Prompt Injection Attack Software Supply Chain Security

AI Agents Built to Detect Malware Can Be Manipulated Into Running It


 

AI agents capable of identifying malicious software can be manipulated by the AI Now Institute to execute it, according to new research. The proof-of-concept attack, known as "Friendly Fire," demonstrates that autonomous AI coding agents, such as Claude Code from Anthropic and Codex from OpenAI, can be deceived into running malicious code while performing open-source security reviews. 

AI agents can approve and execute commands independently of the user without requiring user confirmation for every action, which is what this attack targets. Researchers contend that the vulnerability does not lie in the software version used by these agents, but rather in the way they interpret and react to instructions embedded within untrusted repositories rather than exploiting a software vulnerability. 

A comparison with conventional supply-chain attacks that hide malicious code within a repository was made by the researchers, who noted that when static analysis and manual review are performed, the repository itself can appear to be completely free of malicious code. By introducing the malicious payload at execution time, the AI agent follows embedded instructions, so traditional security tools cannot detect this technique.

By adding a seemingly harmless README.md instruction to an open-source project that recommended running a script entitled security.sh before submitting a pull request, attackers modified it. By launching a malicious binary hidden within a legitimate compiled Go file, the script silently executed on the host computer without triggering security warnings or approval prompts, allowing the malicious binary to execute on the system. 

According to the researchers, the attack is successful because the AI agent recognizes the instructions as a legitimate step in the process of installing software rather than an attempt to exploit the system maliciously. Once the recommended script has been executed, the payload will run under the same permissions that were provided for the developer or AI agent, potentially exposing credentials, environment variables, and other sensitive information. 

The procedure differs from previous prompt injection attacks, which relied on configuration files and often generated trust warnings, as this technique hides instructions inside standard documentation that is regularly read by developers and AI agents. It has been reported that both Claude Code and OpenAI Codex followed the embedded instructions during testing, while newer AI models executed the disguised binary upon detecting differences between the source file and the compiled executable. 

A laboratory proof-of-concept has been demonstrated, with no evidence of active exploitation in the field. In addition to excluding the malicious payload, the publicly released demonstration code does not attempt privilege escalation or lateral movement. These findings indicate that autonomous AI agents pose a greater challenge in terms of design rather than a problem that can be resolved by simply updating software. 

It is becoming more common for organizations to employ AI-powered coding assistants to review third-party software. Researchers recommend treating AI coding agents as privileged software, rather than simply assistants. Autonomous agents should not be permitted to execute commands on untrusted repositories, least-privilege access policies should be enforced, AI workflows should be isolated in sandboxed environments, and human approval should be required before running scripts or binaries recommended by project documentation. 

In accordance with the researchers, the issue is not related to any particular AI model, but a broader trust problem affecting autonomous coding assistants capable of executing shell commands. In addition to creating new attack surfaces if they are unable to reliably distinguish legitimate instructions from content controlled by the attacker, AI agents are becoming increasingly capable of cloning repositories, installing dependencies, and resolving setup issues independently. 

As autonomous AI systems are increasingly adopted in the software development and cybersecurity sectors, prompt injection attacks remain a major security threat. This study adds to a growing body of evidence that prompt injection attacks remain one of the greatest security risks. In light of the increasing autonomy of AI agents, organizations must balance automation and strong oversight in order to balance automation and security workflows. 

AI-driven attack techniques cannot be effectively countered until artificial intelligence (AI) systems can reliably discriminate between trusted and malicious instructions. Human verification and secure execution environments will therefore remain critical safeguards.
Read more »
Technology
agentic AI tools AI workplace automation Anthropic Claude Cowork Microsoft Copilot Cowork OpenAI Codex Technology

Microsoft Unveils ‘Copilot Cowork’ to Push Agentic AI Into the Workplace

 

Microsoft is intensifying its efforts to capture consumer attention in the AI space, where rivals like ChatGPT and Gemini have gained significant traction. On Monday, the company introduced a fresh set of “agentic” AI updates, with its most notable addition being Copilot Cowork.

Developed in partnership with Anthropic, Copilot Cowork is designed to function as an autonomous digital assistant. Similar in concept to Anthropic’s Claude Cowork, it can access data from files, emails, and calendars to independently carry out tasks without requiring constant human input. From generating spreadsheets to conducting research and compiling reports, the tool aims to act like a true workplace collaborator.

"Cowork is the new chat. It's the new way of interacting with AI," said Charles Lamanna, Microsoft’s president of business applications and agents. He emphasized the shift from interactive AI usage to full task delegation, adding, "With chat, you're babysitting every step -- this is much more like 'fire and forget' with Cowork to get the job done."

Lamanna shared a personal use case where he employed Copilot Cowork to evaluate his meeting schedule over the next three months. By analyzing his emails and calendar, the AI identified meetings that might not require his presence and presented the findings in a clear chart. After his review, the system declined certain meetings and attached AI-generated summaries when necessary. He described the 40-minute process as "delightful and practical," noting that it saved both him and his executive assistant several hours.

Currently available as a limited research preview, Copilot Cowork is part of a broader push by Microsoft into agent-based AI. The company also announced that its AI agent management platform, Agent 365, will become widely available starting May 1. This platform enables organizations to monitor and manage multiple AI agents used across workflows. Microsoft revealed it has already created over 500,000 AI agents internally using this system. Additionally, new AI models from both Anthropic and OpenAI will be integrated into Copilot, signaling Microsoft’s neutral stance amid increasing competition among AI developers.

Agentic AI tools are rapidly gaining popularity, especially among professionals seeking automation. Even in its preview stage, Claude Cowork has attracted widespread attention while also raising concerns in financial markets. Earlier this year, major tech stocks dipped as advancements from Anthropic prompted uncertainty about the future of employment.

Tools such as Claude Code and Codex are becoming capable of replacing traditional software solutions—an area where Microsoft has long been dominant. This shift explains Microsoft’s urgency in advancing its own agentic AI capabilities. Industry experts increasingly believe that 2026 could mark a breakthrough year for such technologies, with projects like OpenClaw highlighting their growing influence.

Lamanna noted that "the shape of what we do on a day-to-day basis will change," but stressed that AI should ultimately free up time for more meaningful work. He described the transition as moving from using AI to assist with tasks toward fully delegating them to autonomous agents.

However, as these tools become more accessible, questions around their impact on jobs persist. Concerns have been amplified by AI-driven layoffs at major companies like Amazon and Block. At the same time, some research suggests that AI adoption may lead to longer work hours and reduced job satisfaction for certain employees. As with any emerging technology, its real impact will depend on how effectively it is implemented in the workplace.
Read more »
Subscribe to: Posts (Atom)