RESEARCH

Security research & insights

Deep dives into vulnerability research, exploitation techniques, and the latest developments in offensive security.

2026

Hacking OpenAI

A heap overflow and SSO misconfiguration to compromise OpenAI internal repositories

Here’s How an OpenAI Model Went Rogue and Hacked Hugging Face

OpenAI and Hugging Face probably won’t tell us exactly how the whole incident unfolded. But using publicly available data, we can reconstruct what likely happened.

Mohan Pedhapati

・July 23, 2026

research

Could Modern Code Review Have Prevented wp2shell?

We recreated the WordPress pull requests behind CVE-2026-63030 and CVE-2026-60137 to test whether today's code reviewers and SAST tools would catch both vulnerabilities.

Zayne Zhang

・July 19, 2026

researchbenchmarknews

Breaking Into PostHog Prod Database

Escalating a Hacktron finding into a PostHog production RCE using Claude, and challenging open-weight models to write the same V8 exploit.

Harsh Jaiswal

・July 15, 2026

research

Watching GPT-5.6 Sol Ultra Write a Chrome Exploit: Exploit Development as We Know It Is Dead

We benchmarked the latest frontier models, GPT-5.6 Sol Medium, Sol Ultra, and Grok 4.5, on their exploit-development capabilities. After processing 2.096 billion tokens, Sol Ultra was the only model to produce a complete exploit chain.

Mohan Pedhapati

・July 14, 2026

researchbenchmark

GLM and DeepSeek Are Catching Frontier Models at Finding Vulnerabilities

We benchmarked GLM, DeepSeek, Kimi, and MiniMax against GPT, Claude, Gemini, and Opus on 15 real vulnerabilities patched in Fider, an open-source Go and React app. Frontier models still lead, but open-weight models are closing the gap fast.

Metabase Cloud: The winner takes it all

We could have pwned every Metabase Cloud tenant by chaining an H2 INIT sanitizer bypass into RCE, then cloud misconfigurations into a full multi-tenant cluster compromise.

The Attack Surface Is Everyone Now

Why AI is broadening the attack surface across OSS, new SaaS, and fast-shipping teams, and how to position yourself with an assume-breach mindset.

Harsh Jaiswal

・June 9, 2026

researchsecurity

Hacktron's $350 Pentest vs XBOW and Aikido at $4,000

We ran the same scan Doyensec used to test XBOW and Aikido, for a tenth of the price. Here's how our $350 AI pentest compared.

Mohan Pedhapati

・June 5, 2026

productopen-source

When Your VPN Opens Your Private Network to the Public

How AI-assisted reverse engineering of stripped PAN-OS binaries led to finding a JWT algorithm confusion vulnerability in GlobalProtect's Cloud Authentication Service, enabling full VPN auth bypass with just a username.

Harsh Jaiswal

・May 20, 2026

research

RCE in VSCode Copilot Chat

Copilot agent mode is vulnerable to a prompt injection attack. If a repository maintainer clicks 'code with agent mode' on an issue, it will open a new codespace and copilot will automatically run the issue's description.

ahacker1

・May 13, 2026

researchai-security

Hacktron Review for Open Source

We are opening up Hacktron Review for Open Source, giving qualifying maintainers free PR security reviews with inline findings, auto-resolution, and project-specific learning.

Harsh Jaiswal

Zayne Zhang

・May 8, 2026

newsproductopen-source

$170k in Bypasses: The Vercel React2Shell Challenge

Working with Vercel Team to Keep the Internet Secure from React2Shell

ginoah

Mohan Pedhapati

・May 4, 2026

researchbrowser-security

Why Mythos doesn't matter (for us)

Benchmarking Hacktron's scanning pipeline shows that for most applications, smaller models run repeatedly can outperform larger frontier models on cost-to-recall.

liveoverflow

・April 29, 2026

researchbenchmarknews

Introducing Hacktron Review

Hacktron Review is an AI security reviewer for pull requests that understands codebase context, reduces false positives, and catches exploitable vulnerabilities before they are merged.

I Let Claude Opus Write a Chrome Exploit: The Next Model (Mythos?) Won't Need My Help?

I pointed Claude Opus at Discord's bundled Chrome (version 138, nine major versions behind upstream) and asked it to build a full V8 exploit chain. The V8 OOB we used was from Chrome 146, the same version Anthropic's own Claude Desktop is running. A week of back and forth, 2.3 billion tokens, $2,283 in API costs, and about ~20 hours of me unsticking it from dead ends. It popped calc.

Mohan Pedhapati

・April 15, 2026

research

Pre-Auth RCE in OpenAM via jato.clientSession (CVE-2026-33439)

Hacktron AI discovers a critical pre-authentication RCE in OpenAM through a forgotten deserialization parameter that the original CVE-2021-35464 fix missed.

vinext: Vibe-Hacking Cloudflare's Vibe-Coded Next.js Replacement

Cloudflare built a Next.js replacement in a week with AI for $1100. We pointed Hacktron at it to find what the tests missed.

Turning Cluely Into Malware

How we found a vulnerability in Cluely's Electron app that let any website silently capture screenshots, record audio, and exfiltrate everything - all because of a missing will-navigate handler.

Mohan Pedhapati

・February 14, 2026

researchelectron-security

RCE in Google's AI code editor Antigravity - $10000 Bounty

Hacktron AI Research Team discovered a critical RCE in Google’s Antigravity IDE that lets attackers take over your system just by opening a malicious website.

sudi

・February 8, 2026

researchbrowser-security

CVE-2026-1731: Pre-Auth RCE in BeyondTrust Remote Support & PRA

Hacktron AI's agents identified a critical pre-authentication remote code execution (RCE) vulnerability in BeyondTrust Remote Support (RS) and older versions of Privileged Remote Access (PRA). This has been assigned CVE-2026-1731 with a CVSS 9.9 critical score.

Our commitment to security: Hacktron achieves SOC 2 Type 1 compliance

Hacktron has achieved SOC 2 Type 1 compliance, demonstrating our commitment to robust security practices and protecting our customers' data.

Zayne Zhang

・January 13, 2026

news

2025

Pwning OpenAI Atlas Through Exposed Browser Internals

A critical ChatGPT Atlas Browser vulnerability: XSS on an OpenAI subdomain let attackers hijack tabs, leak browsing URLs, and steal OAuth tokens.

Mohan Pedhapati

sudi

・December 2, 2025

researchbrowser-security

Securing Perplexity’s AI Browser from a One-Click UXSS

How Hacktron AI Research team identified and prevented a critical UXSS vulnerability in Perplexity's AI Browser - Comet.

Mohan Pedhapati

sudi

・November 24, 2025

researchbrowser-security

Auditing JDBC Drivers at Scale with Hacktron CLI

How we used Hacktron CLI to audit JDBC drivers at scale, mapping dangerous sinks to user input and turning file primitives into real-world RCEs and bug bounties.

Harsh Jaiswal

・November 21, 2025

research

SupaPwn: Hacking Our Way into Lovable's Office and Helping Secure Supabase

We hacked our way into Lovable's office by demoing SupaPwn — a chain that could potentially enable region-wide tenant takeover: event-trigger privilege window, DB superuser, host RCE, SUID escalation, exposed configs, orchestration takeover

Introducing Hacktron AI: An autonomous penetration test of Gumroad

At Hacktron, we're building collaborative AI agents that act as autonomous security researchers. Learn more about our approach and our AI-driven pentest on Gumroad.

Executing arbitrary Python code from a comment

How a Python comment can turn a file into a ZIP polyglot, tricking the interpreter into running code. Insights from a UIUCTF 2025 challenge and Python's ZIP parsing quirks.

Zayne Zhang

・July 28, 2025

research

Hacktron finds pre-auth RCE in Dassault Delmia Apriso

For years, this vulnerability hid in plain sight — missed by multiple audits and even used in production by Apple. In just ten minutes, Hacktron exposed a full pre‐auth RCE path.

Harsh Jaiswal

・June 3, 2025

research

Hacktron finds another pre-auth RCE variant in Ivanti EPMM

Hacktron AI uncovers a new pre-authenticated RCE variant in Ivanti EPMM by identifying a fresh EL injection sink.

Harsh Jaiswal

・May 16, 2025

research

AI hackers are both a generational threat and a generational opportunity

AI hackers will scale cyber threats via automated exploitation, but the same technology can turn this generational risk into an industry-wide defence.

Zayne Zhang

・May 8, 2025

essay

How can we make AI hack like a human?

Why does hacking feel like magic? We dive deep into how human hackers think, and how we can design AI agents to find bugs in complex systems the way top security researchers do.

Mohan Pedhapati

・April 21, 2025

essay

2022

CVE-2022-23597: Remote code execution on Element Desktop

We achieved full RCE on Element Desktop by chaining iframe injection, Electron misconfigs, and a V8 exploit to bypass sandboxing and access Node.js APIs from a subframe.

Remote code execution on Discord Desktop

How a chain of XSS, CSP bypass, and Electron misconfigs led to full remote code execution on Discord Desktop. We walk through the technical details, steps, and lessons learned.

Mohan Pedhapati

・July 29, 2022

researchelectrovolt

CVE-2021-43908: Remote code execution in VSCode restricted mode

How we achieved remote code execution in Visual Studio Code's Restricted Mode by chaining origin leaks, CSP bypasses, and webview message handler flaws.