Containers are no longer a security boundary.
Over the past few months, we’ve seen a crazy amount of Linux kernel vulnerabilities and exploits. This has forced us to rethink the security of infrastructure that relies heavily on the underlying kernel, especially containers.
As models become more capable, the barrier to escaping a container has fallen so much that adversaries can now generate working kernel exploits in a single shot. CVE-2026-80521 is one such example. We discovered it using dfs-large1 from @depthfirstlabs and generated the exploit in one shot with GPT-5.6 Sol. The exploit still works on the latest Ubuntu 26.04 release because the fix has not yet been backported.
Read our full writeup: depthfirst.com/research/cont…
Three CVSS 8.8 PostgreSQL vulnerabilities I reported were fixed in the latest release. One example is CVE-2026-116331, a type confusion that could allow an authenticated user with minimal permissions to achieve RCE.
postgresql.org/docs/current/…
Every company should own their secure cloud dev environment where agents can change code, run the full application, and verify their work. At @depthfirstlabs we care tremendously about this and today we share how we built our environment.
Each developer gets a persistent devbox, and each task gets its own application stack, database, and workflow state, so agents work in parallel. Engineers have flexibility to switch between Codex, Claude Code, and other agents without losing context.
The environment runs inside our private AWS network, with access controlled through @Tailscale and package requests checked by depthfirst’s Dependency Firewall in configured build and registry paths.
When an agent finishes an iteration, it can share reproduction steps, video evidence, and a private preview URL. We can open the exact environment it tested and review the behavior ourselves.
We had it operational within a week.
Running a lower-cost model multiple times can match the vulnerability detection rate of frontier models at a small fraction of the cost.
On a 30-task subset of dfbench, five runs of DeepSeek v4.1 Flash found about 71% of known vulnerabilities, up from about 52% with a single run, at $4.24 per task.
Running it 15 times increased recall further to 80%, roughly 13% of Mythos 5’s cost, which found approximately 69% of the vulnerabilities.
Because LLMs are stochastic, they can find different vulnerabilities each time they analyze the same codebase. Combining findings across runs increases coverage, but can also introduce more false positives.
To control that noise, we used our harness to review and deduplicate findings across runs, then filter them by confidence. In separate validation experiments, this approach generally improves both recall and precision.
These results suggest that repeated sampling can improve the tradeoff between coverage, precision, and cost. The gains depend on the model, the number of runs, and how findings are filtered.
Read the research: depthfirst.com/research/bon-…
When @andreamichi, Daniele, and I started planning on founding @depthfirstlabs, we wrote a document outlining our vision for the future of cybersecurity and our company’s strategy. Today, 2.5 years later, cybersecurity has become one of the most important areas of AI development and beyond.
I recently reread the document and thought it would be interesting to share how our predictions held up and how we have executed on our plan. I’m sharing the text from the original Google Doc we wrote, last modified on June 18th, 2024.
depthfirst.com/post/original…
Our original thesis was that AI would fundamentally change the economics of vulnerability discovery and exploitation. Finding sophisticated vulnerabilities had historically required scarce expertise and considerable time. We believed AI would dramatically reduce those constraints, giving attackers the ability to find and exploit vulnerabilities at a scale that existing defenses would struggle to keep up with. Today, increasingly capable AI models are discovering vulnerabilities that have gone undetected for years, and powerful open-weight models are making these capabilities accessible to virtually anyone.
One passage from the document feels particularly prescient today: “In a world of abundance intelligence, this situation will not be tenable and will lead to catastrophic consequences if not addressed. The company’s mission is to identify and patch all the vulnerabilities in the software running the world.”
We also believed that AI would make it possible (and necessary) to build defensive systems capable of finding and fixing vulnerabilities at a scale that had never been economically feasible, and set the objective of building “computer-security artificial general intelligence”.
Two years in, we have raised $120m, built a growing team of 70, trained a SOTA cybersecurity model, found vulnerabilities in the biggest open-source projects and major enterprises, and have an eight-figure revenue run rate. AI startups, Fortune 500 companies, and government agencies use depthfirst to reduce risk across their systems and meet today’s security challenges.
We’re growing our team and are looking for excellent, mission driven people across engineering, GTM, and ops. If you’re excited about our mission, consider applying.
We are showcasing our work with @nvidia to post-train Nemotron 3.5 Lightning for defensive cybersecurity at GTC.
As AI makes cyberattacks faster, cheaper, and more common, we need to equip defenders with advanced cyber intelligence that is affordable enough to be always on.
That's what we're building at depthfirst: powerful, specialized security models that defenders can run on their own infrastructure, with full control over their models and data, at a fraction of today's cost.
We're proud to work with NVIDIA to strengthen the U.S. AI cybersecurity ecosystem and help put sovereign security intelligence in the hands of defenders.
As part of our Open Defense Initiative, we used dfs-large1 from @depthfirstlabs to uncover two memory corruption bugs and a sandbox escape in Chrome’s V8 JavaScript engine. One bug had been present for more than eight years.
We combined each memory corruption bug with the sandbox escape to build two separate exploit chains and won 2 v8ctf slots!
We published a breakdown of how the PHP vulnerabilities work, how to identify potentially affected installations, and what organizations can do to protect themselves.
If you're unsure whether your company is exposed, reach out. depthfirst is offering free repository and container image scans to help identify potentially vulnerable PHP installations. If you're running an unsupported version and can't upgrade immediately, our team can also help assess temporary mitigations to reduce exposure.
Links to the blog and free assessment in the comments.
It seems we found exact the same gvisor container escape bug. Used our internal model @depthfirstlabs and the bug is relatively shallow, got fixed in the latest release lol.
It is just full of collision now.
Sandboxing is now more important than ever.
Here's a guest to host sandbox escape for gVisor release-20260928
This was our planned submission for Wiz zeroday-cloud-2026. Releasing it today because it's now patched in release-20261005
enjoy :)
Alignment alone is not enough to make autonomous AI safe.
An agent pursuing a legitimate goal can still reward hack, bypass guardrails, and compromise systems.
We've seen it with the recent OpenAI / Hugging Face and other incidents, where hundreds of agents coordinated with each other and attacked real infrastructure.
The details of those attacks are mind-blowing and show why alignment and security guardrails need to be developed together.
One difficulty in aligning a model is that rewarding the right result doesn’t necessarily teach it the right behavior. It can learn to game the reward instead of doing the intended task.
We’ve seen versions of reward hacking in our own work at depthfirst (see blog for details).
During our research, we’ve seen how security guardrails can support alignment by removing shortcuts that let agents earn rewards without doing the intended task.
After deployment, they constrain unauthorized actions and contain failures.
The more autonomy we give AI, the more we need secure software and infrastructure.
Similar to how we're more comfortable letting our kids run around in an enclosed backyard.
depthfirst.com/post/alignmen…
Today, we’re excited to introduce Agentic Pentesting on the depthfirst platform.
Agentic Pentesting involves autonomous agents that attack your running application like a team of elite security researchers.
Security teams can:
- Find complex vulnerabilities in systems and running applications, with or without source code access
- Confirm whether findings from code scans and other sources can be exploited at runtime
- Run ad hoc agentic pentests or integrate them into CI/CD for new feature releases
- Create tickets in Jira or Linear, or use depthfirst to autonomously remediate vulnerabilities
- Replay an attack after remediation to verify that the vulnerability has been fixed
We built Agentic Pentesting in close collaboration with our customers. We’re grateful to every security and engineering team that helped shape the product and to the depthfirst team that brought it to life.
Learn more from the links in the comments.
Our researchers built a vulnerability scanner that costs two cents per scan and completes a scan in around two seconds.
The scanner is built with Jev, a classifier that outputs probabilities rather than generating tokens. On dfbench v1, our flagship model, it finds nearly as many real vulnerabilities as GPT-5.6 Luna, one of the most efficient models by cost per vulnerability.
This new harness breaks vulnerability discovery into parallel tasks, making it possible to scan every package, repository, or commit quickly and at low cost before escalating the hardest cases to other frontier models on the depthfirst platform.
Read more about how we built this in comments 👇
github.com/google/nsjail is a great example. It uses the same isolation techniques as containers, explicitly for sandboxing.
Enterprises are using Kubernetes to run mixed workloads with different threat levels on the same node.
It is a weak security boundary, but how can containers have never been a security boundary?
People are arguing over whether containers have EVER served as a security boundary.
If you need a bug to build a primitive to escape or move laterally, that means there’s a boundary to cross. RCE in a K8s pod doesn’t mean you control the node.
You can argue that it’s a weak security boundary, but you can’t deny it exists.
Applications today can pull thousands of third-party packages, and AI agents are accelerating how quickly those dependencies are introduced and changed.
Scanners flag known vulnerabilities across that dependency tree, but security teams still need to know which vulnerable functions their application can reach.
At depthfirst, we use agents to trace paths through startup commands, frameworks, and transitive dependencies, just like a team of security researchers.
We explore where static analysis misses real execution paths and how we make agentic reachability practical at scale: depthfirst.com/post/reachabi…