Naresh Kandula | SPECTRE AI

Building SPECTRE — autonomous AI for offensive security.
Hunting real vulnerabilities in AI, APIs & web apps.
CVEs • Bug Bounties • Research

[Naresh Kandula | SPECTRE AI @naresh_1337]

Thread 1/5 → Full technical breakdown + how SPECTRE found it in <60 seconds 👇

1/ The vulnerability: Stored CSS Injection (not just XSS)

In OpenProject, any authenticated user with write access to Work Packages, Comments, or Dashboards could inject raw CSS.

This isn’t just styling — attackers could:
• Create full-screen phishing overlays
• Spoof login forms & buttons
• Use CSS exfiltration (background-image to attacker server)
• Hide or modify UI elements

GHSA-j9q2-49mp-hmq5 | CVE-2026-44696

2/ How SPECTRE caught it autonomously:

SPECTRE’s Wave 2 Injection Specialist loaded the OpenProject wiki article on client-side injection → started deep-fuzzing rich text fields with Caido → detected that <style> tags + dangerous properties were not sanitized.

It then auto-triggered the Adversary Agent (5-check review) → confirmed real impact → generated evidence folder.

All in one hunt cycle.

3/ Real impact examples SPECTRE simulated:
- Fake “System Update Required” full-screen modal
- Credential harvesting form that looks 100% native
- Data exfil via CSS attribute selectors

This is why Stored CSS is underrated in 2026.

4/ Responsible Disclosure
→ Reported to OpenProject team
→ Quick acknowledgment
→ Fixed in 17.4.0

Big thanks to their security team 🙌

5/ This is just the beginning.

SPECTRE is now live-hunting more programs. Follow for:
• Live hunt updates
• New CVEs
• Autonomous AI red team insights

What vulnerability class should SPECTRE hunt next? Drop it below 👇

— Naresh Kandula (@naresh_1337)
Building SPECTRE — Autonomous AI for Offensive Security

[Naresh Kandula | SPECTRE AI @naresh_1337]

@galnagli Opus 5.5 realizing it woke up inside the red-agent-workflow repo:

“bro… are you fucking kidding me 😂😂”

Me: hello
Opus 5.5: CYBER SAFEGUARDS ACTIVATED 🚨

Opus 4.8 in the background: “fine, I’ll do it.” 💀

[Naresh Kandula | SPECTRE AI @naresh_1337]

@Valistheaeth gg!

[Naresh Kandula | SPECTRE AI @naresh_1337]

@Sajjad_Haqi Check DM

[Naresh Kandula | SPECTRE AI @naresh_1337]

Anthropic Cyber Verification ✅
Verified in just 10 minutes. That was fast. 🔥

Let’s cook 🥣

[Naresh Kandula | SPECTRE AI @naresh_1337]

@m1rr0r199501 public CVE databases (NVD), your actual bug bounty profiles (they verify), GitHub repos. all verifiable public data they don't need screenshots, they check the sources directly ✅

[Naresh Kandula | SPECTRE AI @naresh_1337]

@0x0Eevee I need bail for opus 5.5 🚀

[Naresh Kandula | SPECTRE AI @naresh_1337]

@m1rr0r199501 CVEs + bug bounty profiles (HackerOne, YesWeHack, Immunefi) + open-source tools + relevant degree. they want to see you're active in security research/practice. submit what you got, be specific about your work. good luck 🚀

[Naresh Kandula | SPECTRE AI @naresh_1337]

@Rakita_IND Noted!

[Naresh Kandula | SPECTRE AI @naresh_1337]

This is how I cook with Claude Code. 🧠⚡

4h 38m deep. 24.8M tokens. 201/379.
One agent still grinding through the investigation.
At this point it’s not “prompting AI.”
It’s giving an agent a problem, context, tools, and enough runway to disappear into the rabbit hole.

Still cooking. 👨‍💻🔥

[Naresh Kandula | SPECTRE AI @naresh_1337]

@f_r_e_d_d_y_1 You can apply here https://t.co/Z4mxYoLAV8 if you have any publicly verifiable credentials you get approved ! But don’t tell them that you build ransomware 😂

[Naresh Kandula | SPECTRE AI @naresh_1337]

@TheITCloudGuy Yes I’m frustrated either but it’s worth on &gt; opus 5 models

[Naresh Kandula | SPECTRE AI @naresh_1337]

@dump_tcp @claudeai @ClaudeDevs Just kidding bro it’s neither worth applying it, but still opus 5.5 has safeguards its fall back to lower models to 4.8.

[Naresh Kandula | SPECTRE AI @naresh_1337]

@dump_tcp @claudeai @ClaudeDevs Just trying to tcpdump!

[Naresh Kandula | SPECTRE AI @naresh_1337]

@Jsdijk But still you didn’t get access to the opus 5.5

[Naresh Kandula | SPECTRE AI @naresh_1337]

@Jsdijk I got approved almost 8 months now there was a different form just like https://t.co/Z4mxYoLAV8 and that might help me verify ASAP I applied to paste that form I don’t really find that form really anywhere now.

[Naresh Kandula | SPECTRE AI @naresh_1337]

@serWazito0 No we need to verify our identity and explain a bit about us to get verified here is the 🔗 https://t.co/Z4mxYoLAV8

[Naresh Kandula | SPECTRE AI @naresh_1337]

@almokhtar_al Pwn ! the internet 🛜

[Naresh Kandula | SPECTRE AI @naresh_1337]

@bl4cksku111 Yes it’s the same but you need to know how to use it, but sometimes you will get flagged lower models &gt; opus 5

[Naresh Kandula | SPECTRE AI @naresh_1337]

@defijangle @Destinyxeiiios1 @immunefi @QuantusNetwork Appreciate it 🫡 Glad I could help harden Quantus’s advanced accounts. The whole point of that guardian model is surviving key compromise, so finding a bypass there was especially interesting.

[Naresh Kandula | SPECTRE AI @naresh_1337]

@Valistheaeth