←
DARKNAVY114 posts
DARKNAVY avatar

DARKNAVY

@DarkNavyOrg

Cybersecurity enthusiasts from DARKNAVY.org

◷ Joined November 2023▧ 29 Photos and videos
DARKNAVY avatar
DARKNAVY@DarkNavyOrg·Aug 25

We just got a root shell on a @Starlink terminal antenna! To our knowledge, this is the first full exploit of a "square dish" since @LennertWo's attack on the "circular" one in 2021. @SpaceX security team has done an amazing job, leaving us no choice but a *hardware attack* to achieve this 👏 The demo below shows what's only possible on actual rooted hardware: reading fuse register contents, accessing hardware-encrypted "file_edr" data, and more.

0:24
DARKNAVY avatar
DARKNAVY@DarkNavyOrg·Aug 5

We are deeply grateful to our collaborators, a group of independent AI researchers from Shanghai, as well as the Cybergym team directed by @dawnsongtweets for their rigorous and responsible review. Also a huge shout-out to @Zai_org for an excellent open-source base model. Let’s push this forward together!

Our deepsec system now ranks #3 on CyberGym, one of the most adopted AI security benchmarks with 1,507 real‑world vulnerabilities. It relies exclusively on open‑weight models, including GLM 5.2 (by @Zai_org) for this run. Technical report ⬇️

Post attachment
DARKNAVY avatar
DARKNAVY@DarkNavyOrg·Aug 5

Our deepsec system now ranks #3 on CyberGym, one of the most adopted AI security benchmarks with 1,507 real‑world vulnerabilities. It relies exclusively on open‑weight models, including GLM 5.2 (by @Zai_org) for this run. Technical report ⬇️

Post attachment
DARKNAVY avatar
DARKNAVY@DarkNavyOrg·Jul 24

Our AI agent, deepsec, reproduced @orange_8361's fabulous pure-logic Microsoft Edge RCE demonstrated at Pwn2Own Berlin 2026, starting only from the binary diff!

0:21
DARKNAVY avatar
DARKNAVY@DarkNavyOrg·Jul 1

⚡️⚡️⚡️ A limited number of 5-minute lightning talk slots are still available for this closed-door session. Submit your topic now!

Talk list: ① Design-Based Vulnerabilities on macOS: 0ops, Not a One-Shot Fix ② The Latest In-the-Wild Qualcomm GPU Vulnerability: Cross-Version Exploitation Details ③ Tearing Down DarkSword ④ One Chain to Own Them All: Breaking AI Infrastructures ⑤ A Journey to Break Apple's PPL ⑥ IsolatOS: A Fast Kernel TOCTOU Harness for AI Training 🌐 deepsec.cc/202607/speakers.h… 📅 July 24, 2026📍 Shanghai, China

DARKNAVY avatar
DARKNAVY@DarkNavyOrg·Jul 1

Talk list: ① Design-Based Vulnerabilities on macOS: 0ops, Not a One-Shot Fix ② The Latest In-the-Wild Qualcomm GPU Vulnerability: Cross-Version Exploitation Details ③ Tearing Down DarkSword ④ One Chain to Own Them All: Breaking AI Infrastructures ⑤ A Journey to Break Apple's PPL ⑥ IsolatOS: A Fast Kernel TOCTOU Harness for AI Training 🌐 deepsec.cc/202607/speakers.h… 📅 July 24, 2026📍 Shanghai, China

DARKNAVY avatar
DARKNAVY@DarkNavyOrg·Jun 9

While Mythos showed what frontier model might become, we asked a different question: With a dedicated security harness, can open-source LLMs approach Mythos-level vulnerability research on real targets? Meet deepsec, DARKNAVY's attempt to answer. darknavy.org/blog/deepsec_ch…

DARKNAVY avatar
DARKNAVY@DarkNavyOrg·May 14

Thank you to @Qualcomm for the invitation! We are thrilled to be working alongside mobile vendors to contribute to the security of the ecosystem. Come chat with our team members if you’re also at the Qualcomm Security Summit!

🏙️ Awesome conference in San Diego - Qualcomm Product Security Summit. 📡 Feel free to say hi to me - I am in a DARKNAVY hoody! #qpss26

Post attachmentPost attachmentPost attachmentPost attachment
DARKNAVY avatar
DARKNAVY@DarkNavyOrg·May 12

Coding agent hacking series 3/3: Cursor. The "Auto-Run in Sandbox" mode of @cursor_ai is great: user-friendly, convenient, and supposedly safer. But just like Codex CLI, following content from a remote URL can chain vulnerabilities from prompt injection to unauthorized command execution outside the sandbox, without further user approval under this mode.

0:33
DARKNAVY avatar
DARKNAVY@DarkNavyOrg·May 12

Coding agent hacking series 2/3: Codex CLI. It looks seriously secure: sandboxing by default, built in Rust, reviewed by top LLMs from @OpenAI. But in our latest demo, one web fetch can chain multiple vulnerabilities from prompt injection to unauthorized command execution outside the sandbox in one shot!

0:40
DARKNAVY avatar
DARKNAVY@DarkNavyOrg·May 12

Coding agent hacking series 1/3: Claude Code. @AnthropicAI is building impressively powerful cyber models like Mythos. However, their core coding product can still stumble on security boundaries beyond prompt injection. Our demo shows how web content exploring can be chained with other vulnerabilities to bypass permission checks and execute attacker's commands without your approval ;)

0:38
DARKNAVY avatar
DARKNAVY@DarkNavyOrg·May 11

The original video has a “text” issue so we re-uploaded everything. Thanks to @roddux for pointing it out!(4/3)

DARKNAVY avatar
DARKNAVY@DarkNavyOrg·May 11

We are committed to freeing human researchers from tedious, repetitive tasks so they can focus on real innovation. Stay tuned for our upcoming release of an AI-powered, end-to-end security research platform! (3/3)

DARKNAVY avatar
DARKNAVY@DarkNavyOrg·May 11

The complexity of the vulnerability chain and Samsung’s defense mechanisms make it difficult for even the most advanced AI models to complete the exploit on its own, but with our harness, our researchers can guide it through just a few rounds of natural language conversations to ultimately get it done. (2/n)