Delphos Labs (@DelphosLabs) header
Delphos Labs (@DelphosLabs) profile picture

AI for compiled code – no source needed. Analyze malware & third-party software. Unlock natural language summaries, component manifests, security mitigations.

Joined July 2024
Photo by Delphos Labs (@DelphosLabs)Photo by Delphos Labs (@DelphosLabs)Photo by Delphos Labs (@DelphosLabs)
Delphos Labs retweeted
Delphos threat analysis: KarstoRAT is a Windows RAT that spies through the webcam, microphone, and keyboard. With one operator command, it removes its persistence and deletes itself via a hidden batch script. While it runs, it also: - Steals tokens from Discord, Chrome, Brave, and Opera - Gives the operator a remote shell, screenshots, and clipboard monitoring - Elevates itself to admin through a fodhelper UAC bypass, without a standard UAC prompt Lantern, our platform at @DelphosLabs for analyzing software artifacts without source code, found all of this in the sample alone. Have a suspicious Windows file you think will stump us? Reach out or request access: delphoslabs.com/request-acce… Full report, with scope, behavioral sequence, and evidence: delphoslabs.com/report/karst…
3
4
123
Delphos Labs retweeted
The @LastPass TIME team and Delphos identified and disrupted a multi-stage malware delivery campaign impersonating at least 40 companies on GitHub. The payload delivered a Microsoft-attested kernel driver that terminated 145 antivirus and EDR processes from kernel mode. It scored 0 out of 72 on VirusTotal as of August 20. The driver is a renamed copy of CcProtect.sys, which LOLDrivers already catalogues as a BYOVD process killer, with public PoC code. Same product string, version, and submitter. The operators changed the filename to Alinubx.sys. That was the evasion. Detections keyed to the known name did not match, and renaming a file changes its bytes, so the SHA256 moved with it. A blocklist of known bad hashes does not have this one. Same driver, same version, same submitter, one new filename. That was enough. Joint writeup with @LastPass. Hashes, hunt queries, and the full CcProtect comparison are in the post. delphoslabs.com/blog/alinubx…
2
3
13
783
Delphos Labs retweeted
One of our security researchers at Delphos Labs, Kamil Leoniak, triggered kernel memory corruption on Windows using a custom USB device and a user-mode stress PoC. Microsoft assigned CVE-2026-50321, classified it as a Windows USB Driver elevation of privilege, and shipped the fix on July 14. The bug is that the freed object reference stayed reachable long enough for another path to use it. WinUSB_FreePipe frees a pipe context and only nulls the shared pointer afterward. In that window the interface pipe array still points at freed memory, and device-control IOCTLs dispatch on a WDF parallel queue, so nothing serializes access to that array. 16 threads were enough. 12 calling pipe policy APIs, 4 toggling alternate settings. A policy call read the stale pointer and took a spinlock inside a freed 0xF0-byte NonPagedPool object. Without Special Pool that allocation stays mapped and reclaimable. With it, the machine bugchecks. Full writeup on the Delphos Labs blog, with the PoC on GitHub. delphoslabs.com/blog/cve-202…
5
1
19
1,541
Delphos Labs retweeted
60+ AV engines called this file clean. The Delphos Labs platform called it ValleyRAT in five minutes, a full day before the first signature landed. A freshly uploaded Windows executable, zero detections. The filename was the first tell: a Chinese-language lure translating roughly to "click here to visit the official Crab website." From the binary alone, no sandbox detonation, no prior signature, no threat intel feed, we identified it as a Chinese RAT stager. Static analysis and emulation surfaced networking imports, process creation, crypto primitives, and anti-analysis checks. The behavioral chain confirmed it. The stager pulls a ZIP from an AWS S3 bucket, extracts it, and drops a randomized-name second stage. String extraction returned the C2, a domain tied to carding forums. The vendors caught up 24 hours later, long enough for the stager to run. IOCs below. delphoslabs.com/blog/39d4237…
4
10
556
The Hugging Face incident post-mortem is out, from the CSA CISO Community with SANS, [un]prompted, RSAC, Knostic, and FIRST. Our CISO Joe Szczerba was one of the contributing authors. Guidance is sequenced into what to do this week, this month, and this quarter.
3
232
On July 30th we'll be at the AI Cyber Alliance Meetup in Tysons, VA. Our CISO Joe Szczerba, ex-FBI Cyber Division Senior Exec who defended critical infrastructure from nation-state threats, speaks with @epowell101, CEO of DeepTempo. Come say hi 👉 luma.com/aicyber-y9ry?utm_so…
2
2
194
DirtyCBC / DirtyDecrypt (CVE-2026-31635?) Write-up: github.com/Delphos-Labs/disc… Another exploit: github.com/v12-security/pocs…
1
7
775
Delphos Labs retweeted
A credential stealer took the name of libpsl-5.dll, a Windows library of approximately 200KB, and inflated itself to 55MB. Zero detections across 60+ AV engines when we analyzed it. The Delphos Labs platform pulled it apart on day one; the automated run took 8 minutes, with no human intervention. The vendors did not catch up for six days. The bloated size is the primary evasion mechanism. The real code ends at about 3.3MB. The other 52MB is a single four-byte pattern repeated across the entire overlay, sitting outside every PE section and outside the image the Windows loader maps. The loader never reads it. It executes nothing. Its only job is to push the file past the size most engines and sandboxes scan. No signatures. Detection came from the size mismatch, plus an import table pairing process injection with network enumeration, nothing a suffix-list library would do. Its code overlaps several known stealer families. delphoslabs.com/blog/39d4237…
5
3
419
Delphos Labs retweeted
One of our security researchers demonstrated a local root shell on Linux using a page-cache poisoning primitive in AF_RXRPC’s RxGK path. We call it DirtyCBC: a sibling to DirtyFrag in the broader CopyFail / DirtyFrag / Fragnesia family. The issue is fixed on mainline. The candidate path was surfaced through Delphos’s agentic analysis workflow, then manually verified and exploited end to end. AES-256 was not broken. It just wasn’t the boundary that mattered. RxGK decrypted data in place before authentication completed. Under the right conditions, that write could land in the page cache. The HMAC check still failed and the connection was aborted, but the page-cache mutation had already happened. Two RESPONSE packets were enough to place a tiny ELF into the cached first page of a readable SUID-root binary. The file on disk stayed unchanged. The next exec produced a root shell. Full writeup and PoC on the Delphos Labs GitHub. delphoslabs.com/blog/3614237…
5
67
5
288
63,226
Delphos Labs retweeted
If you hear “agentic” three times on the RSAC floor, a startup gets funded. Just kidding. What is true? Security is moving from hype to implementation. Filmed live at @OneRSAC for The Phish Bowl, featuring @DelphosLabs Co-founder, @caleb_fenton.🎙️ youtube.com/watch?v=dwo4UWjk…
1
1
624
Stay tuned.
Our automated reverse engineering systems are starting to find vulnerabilities. First of many. All black box. Only the compiled binary. Identified using code reasoning. Responsible disclosure underway.
177
🎃 Happy Halloween from the Delphos Labs team! “Ghosts in the Machine” 👻 The only things scarier than hackers? Our Halloween costumes. 💀 Even people who work in the shadows of code need a frightfully good time. No incidents were reported, just excellent costumes.
145
Delphos Labs retweeted
My employer, @DelphosLabs, is surveying the RE community to optimize the impact of our future development. It's only open for one more week. Last chance to participate! We'd love to learn more about you, your needs, or feedback on the public beta! Try it! docs.google.com/forms/d/e/1F…
1
4
5
1,929
What keeps CISOs up at night? AI risk. Vulnerability management. Staying resilient when the pressure’s real. New threats move fast. Regulators are louder. Consumers are losing patience. Boards want answers. And AI makes it easier than ever for anyone to write an exploit.
2
1
4
260
Security leaders aren’t just fighting attackers, they’re navigating shrinking timelines, mounting expectations, and unclear signals. The risks are evolving. Are your defenses? Software keeps shipping. Resilience can’t be an afterthought.
1
132
We unpacked all this with CISOs & Security Leaders at our @BlackHatEvents breakfast. Big thanks to @AverlonAI for co-hosting. The best convos weren’t on stage — and we’re just getting started. #BlackHat2025 #BlackHat #Cybersecurity
110
We flagged this binary as Conti-style ransomware before checking threat intel. All automated. No unpacking. No signatures. Just behavior: Threaded loader. AES/RSA encryption. Dynamic API calls. It screamed “Conti” before we did. Patterns > signatures. delphoslabs.com/uploads/2b4e…
1
3
312