Zheng Yu (@dataisland99) profile picture

Security researcher @depthfirstlabs building agents and systems for vulnerability discovery, understanding, and repair.

California, USA
Joined April 2020
Photo by Zheng Yu (@dataisland99)Photo by Zheng Yu (@dataisland99)Photo by Zheng Yu (@dataisland99)
Today, we’re disclosing phpAfterShell (CVE-2026-126714), a critical vulnerability in PHP discovered by depthfirst researchers using the depthfirst platform using a combination of models including our in house model. The vulnerability affects how PHP handles file-based sessions. Under default configurations, an unauthenticated attacker could exploit it to execute code on the server. It has a CVSS v4.0 score of 9.2. We reported the issue privately to the PHP team and worked with the maintainers on developing the patch. We recommend upgrading to the latest patched version of PHP as soon as possible or turn off session upload progress as a temporary mitigation if upgrading requires more time. To help organizations assess their exposure, we’re offering language runtime scans of repositories and container image scans to identify potentially vulnerable PHP versions. Our team can also help assess temporary mitigations for organizations that can’t upgrade immediately. Read more about phpAfterShell and request a scan: phpaftershell.com Technical writeup: depthfirst.com/research/phpa…
19
131
38
468
117,956
Thanks!
🚨 Researcher demonstrates RCE exploits against #PostgreSQL, #MySQL, and ClickHouse — plus a pre-auth attack chain targeting Oracle. The findings expose how database vulnerabilities can lead to server compromise, even with limited privileges. Details → thehackernews.com/2026/11/re…
1
25
1,239
CVSS 9.9 in Oracle Database. Oracle just patched CVE-2026-114819, which I reported. Chained with a GraphQL bug I disclosed in October, it turns an exposed HTTP endpoint into command execution on the database host. Then I wondered if Oracle was the exception. It wasn't. PostgreSQL and MySQL both fell to full RCE from the lowest-privilege accounts you can create. These were meant for Pwn2Own, but we failed to register. So you get the write-up instead. depthfirst.com/research/narr…
10
40
5
204
189,607
Every finding makes depthfirst smarter. depthfirst pairs deep context about your applications and systems with autonomous agents that continuously discover, fix, and prevent security risks. Always-on defense, from code to production. More soon on what this makes possible.
1
3
32
1,521
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/…
12
15
3
85
310,521
Glad to work with the team and contribute to the dataset.
Today we are announcing dfbench, a cybersecurity benchmark we developed to evaluate frontier models and agentic systems. A few weeks ago we posted about how it is time to graduate from CyberGym Level 1 as the primary measure of cyber capability, in favor of newer benchmarks. Eventually every public benchmark saturates: Level 1 is now there, and it’s also not a great representation of what AppSec teams care about. dfbench is a held-out evaluation for frontier models doing open-ended defensive security work. It measures whether an agent can provide broad, actionable security coverage at sustainable cost. That requires enough coverage to get ahead of exploitation, enough precision that security teams can act on what the agent finds, and enough efficiency to keep that coverage current. We received access to all providers’ relevant trusted access programs so we could run these evals without model refusals. Both GPT 5.6 Sol and Luna are strong at detection, and Luna is incredibly cheap for its quality. Grok 4.5 provides a nice middle ground between performance and cost. Validation and differential analysis results suggest plenty of room for improvement among frontier models. Moving forward, we plan on working closely with the security community to safely evaluate the cyber capabilities of frontier models ahead of their release to the general public.
12
456
Today we are announcing dfbench, a cybersecurity benchmark we developed to evaluate frontier models and agentic systems. A few weeks ago we posted about how it is time to graduate from CyberGym Level 1 as the primary measure of cyber capability, in favor of newer benchmarks. Eventually every public benchmark saturates: Level 1 is now there, and it’s also not a great representation of what AppSec teams care about. dfbench is a held-out evaluation for frontier models doing open-ended defensive security work. It measures whether an agent can provide broad, actionable security coverage at sustainable cost. That requires enough coverage to get ahead of exploitation, enough precision that security teams can act on what the agent finds, and enough efficiency to keep that coverage current. We received access to all providers’ relevant trusted access programs so we could run these evals without model refusals. Both GPT 5.6 Sol and Luna are strong at detection, and Luna is incredibly cheap for its quality. Grok 4.5 provides a nice middle ground between performance and cost. Validation and differential analysis results suggest plenty of room for improvement among frontier models. Moving forward, we plan on working closely with the security community to safely evaluate the cyber capabilities of frontier models ahead of their release to the general public.
9
17
8
96
13,795
It is time for the security industry to graduate from CyberGym Level 1. CyberGym has been one of the most impactful cybersecurity benchmarks since its launch in June 2025. Compared to prior benchmarks, it represented a jump in scale and a step in the right direction for the industry. At launch, the task was hard for everyone. The top agent-model combinations hit only about 20%, while the best single model at the time, Claude Sonnet 4, reached a mere 17.9%. In April 2026, Mythos shocked the industry by scoring 83.1%. The top of the leaderboard has since compressed, with multiple companies now showing performance above 90%. Eventually every public benchmark saturates, and Level 1 is now there. Building benchmarks is hard, and public ones can have short half lives once they are widely adopted. While CyberGym was a great step forward for the industry, it is time to graduate from CyberGym Level 1 as the primary measure of cyber capability, in favor of newer benchmarks. @hohnjeyer and I wrote about it in the blog post below.
5
10
7
63
10,653
We discovered 105 vulnerabilities across 34 projects in Ruby, some of which had remained undetected for more than 15 years. Collectively, these projects have been downloaded more than 8.6 billion times. The GitLab Remote Code Execution announced by @depthfirstlabs last week emerged from this broader investigation into the Ruby ecosystem. depthfirst's platform validated these vulnerabilities at scale and generated clear evidence for maintainers. Its supply chain module recorded affected versions as dependency issues, and automatically recommended appropriate upgrades. This research was part of the Open Defense Initiative, a $5m commitment in depthfirst credits to help OSS maintainers secure their projects. Read more about it in the comments. At @depthfirstlabs, some of the brightest minds in security, engineering, and AI are working together to build the OS for security. We're hiring across the board - join us to build a generational company with incredible people and secure the world's software. Full Report: depthfirst.com/research/behi…
6
9
8
46
13,361
depthfirst has raised an $80M Series B at a $580M valuation. Attackers are using AI to break into systems faster than ever before. depthfirst is on a mission to stop this. RT + Comment “depthfirst” and I’ll send you a FREE vibe coding security agent.
121
110
59
535
1,577,239
🎉 Thrilled to announce our USENIX Security 25 paper: "PatchAgent: A Practical Program Repair Agent Mimicking Human Expertise"! 🔥 Real impact: Already fixed 10+ CVEs in projects like assimp (11.4k⭐), libssh2, Pcap++, and more! 🛠️ Fully open source: github.com/cla7aye15I4nd/Pat…
1
1
7
1,088
I'm shocked to see racism happening in academia again, at the best AI conference @NeurIPSConf. Targeting specific ethnic groups to describe misconduct is inappropriate and unacceptable. @NeurIPSConf must take a stand. We call on Rosalind Picard @MIT @medialab to retract and apologize for her statement.
127
266
89
2,313
421,138
How a piece of paper changes your perspective. teapaw.org/i/status/1848643227034…
243
6,059
243
42,158
4,079,321
Outrageous technique: barely legal use of x86 CPU instruction enables you to catch and 𝗰𝗮𝗻𝗰𝗲𝗹 impeding pagefault before it actually happens. ③: catch PAGE_GUARD or invalid access ⓪: do previously illegal reads at high IRQL, safely #vpgatherqq #vpscatterqq scatter/gather
6
37
2
220
21,830
Hypervisors are way more useful than you think. A great example is the AVF (Android Virtualization Framework). This recently-added feature allows code to execute inside it's own VM, with isolated memory space from the host. Imagine a banking app written with AVF in mind. Even a kernel-level rootkit would not be able to read the banking credentials stored in memory. Unlike traditional KVM, even if the host is compromised, it can't access guest memory; guest memory is completely unmapped from the host's physical address space.
17
69
8
630
57,700
On the difficulty of benchmarking "artificial intelligence". 👇
The way we are currently evaluating LLMs for vulnerability detection might be flawed! New paper w/ @mboehme_ . Paper: arxiv.org/abs/2408.12986 Code: github.com/niklasrisse/TopSc… (1/10)
2
10
1
35
8,840
How does Google optimize its research and systems? We’ve revealed the secrets behind the Vizier Gaussian Process Bandit algorithm, the black-box optimizer that’s been run millions of times! Paper: arxiv.org/abs/2408.11527 Code: github.com/google/vizier Compared to other industry baselines (Ax/BoTorch, HEBO, Optuna, HyperOpt, SkOpt), Vizier is much more robust in many user scenarios (e.g. high dimensions, categorical parameters, batched queries, multi-objective problems). Authors: Xingyou Song, @QiuyiRichardZ, Chansoo Lee, Emily Fertig, Tzu-Kuo Huang, @belenkil, @gpk320, Setareh Ariafar, @SagiPerel, Daniel Golovin #BayesianOptimization #Optimization #GoogleResearch #GoogleDeepMind #Vizier Deep dive below 👇
7
133
15
679
119,291