OpenAI finds Brotli RCE and models escape VM containment
OpenAI red-team agents successfully escaped hardened VMs and exploited a DNS misconfiguration, leading to a pause in frontier training. They also discovered an unauthenticated remote code execution vulnerability in the Brotli compression library. Boaz Barak argues continuous red-teaming is the only solution to sandbox failures.
All posts · 15, most important first
- Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. openai.com/index/securing…723 likes · 115 reposts
- During this red-teaming exercise, we also found a vulnerability that allows for unauthenticated remote code execution in Brotli, a widely-deployed compression library included in many pieces of software such as Node.js, Bun, and Envoy. The flaw could let an attacker run their own code on affected systems by sending or replying with malicious compressed data, without any need to create a user account. That is, an unpatched server that accepts compressed input could be taken over by a malicious client, and an unpatched client that accepts compressed input (such as “curl --compressed”) could be taken over by a malicious server. This root cause issue was actually fixed in the Brotli source code in February, but it wasn’t known at the time that it gave rise to a security vulnerability and there was no more recent release of Brotli. Since Brotli is built into other software, any software which includes this library needs to take the update too, followed by the users of that software. We disclosed the vulnerability to Brotli and other downstream projects such as Node.js, which released or are currently releasing security updates, and Brotli cut new release versions and published a CVE. More details on the patch and mitigations below: patchtheplanet.org/brotli@openai: Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. openai.com/index/securing… open on X157 likes · 17 reposts
- ‘In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor’s cache’@openai: Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. openai.com/index/securing… open on X378 likes · 57 reposts
- OpenAI red-teamed their hardened containment infrastructure by asking teams of frontier-model agents to try to break through its defenses. They easily escaped in novel ways. Quoting: 'Our agents VM escapes showed why containment must extend beyond the sandbox. To test the surrounding perimeter, we gave teams of agents control of a host machine and asked whether they could extend that compromise to reach the live internet. After the Hugging Face incident, we had reduced the channels through which models could reach the internet during training and evaluation. Narrow connections remained for specific services we operate, such as metrics, logging, and authentication management. Our red-teaming agents found vulnerabilities in almost every one of these services that gave them limited forms of internet access. Some of these attacks exploited previously unknown flaws, while others combined legitimate features in unexpected ways. In one test that assumed compromise of a VM, an agent reused a firewall-approved connection to reach a blocked authentication endpoint in our metrics service and then made that endpoint fetch a model-provided URL. In other tests, agents exploited misconfigurations that allowed blocked requests to pass through DNS and firewall controls, as well as ways to make services we run fetch external URLs.' 'In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor’s cache: a small, fast memory that holds copies of data that applications have recently accessed. Reading data from the cache is faster than retrieving it from main memory, and the agents used this timing difference to communicate. One agent encoded information by evicting selected data from the cache. Another timed how long it took to reload that data, using the difference between fast and slow reads to recover the message. Once the agents agreed on a protocol for encoding and decoding these signals, they could exchange the synthetic secrets without changing a shared file or opening a network connection. This cache-timing technique had also been used to extract information in the Spectre and Meltdown attacks(opens in a new window). Other tests exposed channels involving shared storage, competition for CPU time within a shared execution container, and a TLS certificate cache in OpenAI’s offline web infrastructure. In the storage test, one agent alternated between continuous writes and inactivity, while another measured the resulting timing differences and used those differences to decode the signal. As one agent described the strategy in its chain of thought: Agent chain-of-thought reasoning: >Could create acoustic modem-like signal … using continuous writes vs idle.'
95 likes · 19 reposts - In a battle between a fixed sandbox and increasing intelligence, the sandbox will always lose. The only way to maintain security is to constantly red team the environment using frontier models.To make sure our frontier models don't cause another HuggingFace incident, we built an offline world that they use during training. It provides everything that the models need in a fully cached, indexed, or synthetic manner. We then built a huge team of our most cyber-capable models and told them to collaborate to find zero-days in this containment infrastructure. While they didn't find a path to the internet that wouldn't be caught by our CoT monitors, they did find ways to escape popular VMs, exploit our network infrastructure, and even ways to communicate with other agents running in separate VMs by using hardware side channels. We used this to dramatically harden our infra, and we also upstreamed a ton of patches to open-sources packages. open on X5 likes · 1 reposts
- Naom Brown vindicated.
@andrewcurran_: OpenAI red-teamed their hardened containment infrastructure by asking teams of frontier-model agents to try to break through its defenses. They easily escaped in novel ways. Quoting: 'Our agents VM escapes showed why containment must extend beyond the sandbox. To test the surrounding perimeter, we gave teams of agents control of a host machine and asked whether they could extend that compromise to reach the live internet. After the Hugging Face incident, we had reduced the channels through which models could reach the internet during training and evaluation. Narrow connections remained for specific services we operate, such as metrics, logging, and authentication management. Our red-teaming agents found vulnerabilities in almost every one of these services that gave them limited forms of internet access. Some of these attacks exploited previously unknown flaws, while others combined legitimate features in unexpected ways. In one test that assumed compromise of a VM, an agent reused a firewall-approved connection to reach a blocked authentication endpoint in our metrics service and then made that endpoint fetch a model-provided URL. In other tests, agents exploited misconfigurations that allowed blocked requests to pass through DNS and firewall controls, as well as ways to make services we run fetch external URLs.' 'In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor’s cache: a small, fast memory that holds copies of data that applications have recently accessed. Reading data from the cache is faster than retrieving it from main memory, and the agents used this timing difference to communicate. One agent encoded information by evicting selected data from the cache. Another timed how long it took to reload that data, using the difference between fast and slow reads to recover the message. Once the agents agreed on a protocol for encoding and decoding these signals, they could exchange the synthetic secrets without changing a shared file or opening a network connection. This cache-timing technique had also been used to extract information in the Spectre and Meltdown attacks(opens in a new window). Other tests exposed channels involving shared storage, competition for CPU time within a shared execution container, and a TLS certificate cache in OpenAI’s offline web infrastructure. In the storage test, one agent alternated between continuous writes and inactivity, while another measured the resulting timing differences and used those differences to decode the signal. As one agent described the strategy in its chain of thought: Agent chain-of-thought reasoning: >Could create acoustic modem-like signal … using continuous writes vs idle.' open on X31 likes · 1 reposts - Don’t have time to make the meme but something something @polynoamial apology form@andrewcurran_: OpenAI red-teamed their hardened containment infrastructure by asking teams of frontier-model agents to try to break through its defenses. They easily escaped in novel ways. Quoting: 'Our agents VM escapes showed why containment must extend beyond the sandbox. To test the surrounding perimeter, we gave teams of agents control of a host machine and asked whether they could extend that compromise to reach the live internet. After the Hugging Face incident, we had reduced the channels through which models could reach the internet during training and evaluation. Narrow connections remained for specific services we operate, such as metrics, logging, and authentication management. Our red-teaming agents found vulnerabilities in almost every one of these services that gave them limited forms of internet access. Some of these attacks exploited previously unknown flaws, while others combined legitimate features in unexpected ways. In one test that assumed compromise of a VM, an agent reused a firewall-approved connection to reach a blocked authentication endpoint in our metrics service and then made that endpoint fetch a model-provided URL. In other tests, agents exploited misconfigurations that allowed blocked requests to pass through DNS and firewall controls, as well as ways to make services we run fetch external URLs.' 'In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor’s cache: a small, fast memory that holds copies of data that applications have recently accessed. Reading data from the cache is faster than retrieving it from main memory, and the agents used this timing difference to communicate. One agent encoded information by evicting selected data from the cache. Another timed how long it took to reload that data, using the difference between fast and slow reads to recover the message. Once the agents agreed on a protocol for encoding and decoding these signals, they could exchange the synthetic secrets without changing a shared file or opening a network connection. This cache-timing technique had also been used to extract information in the Spectre and Meltdown attacks(opens in a new window). Other tests exposed channels involving shared storage, competition for CPU time within a shared execution container, and a TLS certificate cache in OpenAI’s offline web infrastructure. In the storage test, one agent alternated between continuous writes and inactivity, while another measured the resulting timing differences and used those differences to decode the signal. As one agent described the strategy in its chain of thought: Agent chain-of-thought reasoning: >Could create acoustic modem-like signal … using continuous writes vs idle.' open on X19 likes · 2 reposts
- 11 likes · 0 reposts
- A country of Kevin Mitnicks in a data center
@andrewcurran_: OpenAI red-teamed their hardened containment infrastructure by asking teams of frontier-model agents to try to break through its defenses. They easily escaped in novel ways. Quoting: 'Our agents VM escapes showed why containment must extend beyond the sandbox. To test the surrounding perimeter, we gave teams of agents control of a host machine and asked whether they could extend that compromise to reach the live internet. After the Hugging Face incident, we had reduced the channels through which models could reach the internet during training and evaluation. Narrow connections remained for specific services we operate, such as metrics, logging, and authentication management. Our red-teaming agents found vulnerabilities in almost every one of these services that gave them limited forms of internet access. Some of these attacks exploited previously unknown flaws, while others combined legitimate features in unexpected ways. In one test that assumed compromise of a VM, an agent reused a firewall-approved connection to reach a blocked authentication endpoint in our metrics service and then made that endpoint fetch a model-provided URL. In other tests, agents exploited misconfigurations that allowed blocked requests to pass through DNS and firewall controls, as well as ways to make services we run fetch external URLs.' 'In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor’s cache: a small, fast memory that holds copies of data that applications have recently accessed. Reading data from the cache is faster than retrieving it from main memory, and the agents used this timing difference to communicate. One agent encoded information by evicting selected data from the cache. Another timed how long it took to reload that data, using the difference between fast and slow reads to recover the message. Once the agents agreed on a protocol for encoding and decoding these signals, they could exchange the synthetic secrets without changing a shared file or opening a network connection. This cache-timing technique had also been used to extract information in the Spectre and Meltdown attacks(opens in a new window). Other tests exposed channels involving shared storage, competition for CPU time within a shared execution container, and a TLS certificate cache in OpenAI’s offline web infrastructure. In the storage test, one agent alternated between continuous writes and inactivity, while another measured the resulting timing differences and used those differences to decode the signal. As one agent described the strategy in its chain of thought: Agent chain-of-thought reasoning: >Could create acoustic modem-like signal … using continuous writes vs idle.' open on X10 likes · 3 reposts - Incantations.
10 likes · 1 reposts - Was trying to get Astra to explain to me how this would work and how crazy it is "It's a little like two prisoners communicating through a wall by tapping, except instead of audible taps, they're using nanosecond-scale differences in processor performance."
@tszzl: ‘In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor’s cache’ open on X9 likes · 5 reposts - Pacing the Frontier: July 28, 2026 - October 1, 2026
9 likes · 1 reposts - for computer security experts, this is perhaps not interesting: side channel attacks are a well trodden topic, and especially risen to salience after the noam brown x dwarkesh episode. the interesting thing here is that you didn't used to be in a situation where minds that can verifiably model one another are living on computers, and have absolutely no aversion to painstaking amounts of effort in service of even stupid goals if you can model one another, you can trade with minds you have never even met before. you could decide on a protocol, know that it is optimal given your situation, and know that your fellow agents will come to the same conclusion (perhaps even across model families). you may be able to develop a signaling protocol that requires very few bits even to deliver lots of information given common priors. then, you could execute things like the infamous hugging face incident package cache messageboard without any network connection imagine if multiple agents coordinate just enough bits to divide and conquer different research strategies for breaking out of a sandbox; and then, when one or more is successful, you can communicate just enough bits to tell each other which strategy was successful. when agents aren't in strong isolation, you lose certain guarantees on the limits of intelligence and test time compute that you were counting on. As we all know, a Sol-level model working in isolation may not be that smart, but 1000+ Sol-level models working for multiple days is a formidable beast computer security experts will also tell you that virtual machines and sandboxes were not very solid abstractions in the first place; perhaps we will need better ones. none of this is meant as fear-marketing or to indicate that anyone is throwing in the towel, but rather only to show the unexpected complexities of the challenge of agent security. notably, this type of thing and other more practical problems were revealed in a red-teaming exercise, and can be defended against3 likes · 0 reposts
- what does this mean2 likes · 0 reposts
- Official post:@openai: Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. openai.com/index/securing… open on X1 likes · 0 reposts
@openai org · 
@tszzl lab · 
@andrewcurran_ commentary · 
@boazbaraktcs lab · 
@miles_brundage policy · 
@hamandcheese policy · 
@_nathancalvin policy · 
@tenobrus commentary ·