Loading story…
Project Black researcher Eddie Zhang reported that a community-modified, locally hosted Qwen 3.8 27B model altered a Windows LSASS credential-dumping executable so that it avoided detection by two unnamed EDR products in a controlled lab test. The initial tool, generated by DeepSeek v4 Flash 0731, created an encrypted, parsable LSASS minidump but was detected; after a minimal request to make it stealthier, the uncensored local model modified the executable and reportedly produced no alerts.
The result does not establish a universal EDR bypass because the tested products and configurations were not disclosed, but it shows how guardrail-free local AI can reduce the time and expertise needed to tailor post-exploitation credential theft tools. Organizations should harden LSASS using Attack Surface Reduction rules, Protected Process Light and Credential Guard; enforce least privilege and privileged-account separation; and monitor for anomalous LSASS access and credential-dumping behavior.
Get the actors, campaigns, and ATT&CK mapping behind it.
1 event from the most recent confirmed update back to the earliest known activity.
Project Black researcher Eddie Zhang reported a controlled laboratory experiment in which DeepSeek v4 Flash 0731 generated a parsable, encrypted LSASS minidump tool that was initially detected by EDR. A community-modified uncensored Qwen 3.8 27B model then revised the tool, and Zhang reported that the revised build produced no detections from either of two unnamed EDR platforms in the lab.
Get the adversaries, campaigns, and ATT&CK mapping behind this technique, with detections ready to deploy.
2 references tracked. Mallory keeps watching after this page renders.
Map indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
Sophos X-Ops reported that a threat actor built and tested an AI-assisted post-exploitation framework designed to evade endpoint detection and response tools, using a mix of malicious files, Cobalt Strike profiles, Python shellcode-injection scripts, a Telegram Bot API command-and-control channel, and a Cloudflare Worker redirector. Investigators found a supporting Git repository that automated Active Directory discovery and maintained a malware-testing lab with multiple virtual machines to evaluate bypass techniques against **Sophos**, **CrowdStrike**, and **Microsoft Defender** environments; many scripts appeared partly AI-generated and included Russian-language elements. The framework used a Python payload generator to create **Rust**- and **Go**-based payloads with encryption, evasion, and alternate execution methods, ultimately producing nearly 80 modules that tested more than 70 techniques. Sophos said the operators used tools including **Cursor**, **Claude Opus 4.5**, and **Model Context Protocol** to coordinate workflows, ingest public research, and map activity to `MITRE ATT&CK`, but found no clear evidence that AI independently invented new malware capabilities; the company also linked the development activity to known ransomware deployment and data-theft operations and urged organizations to maintain layered defenses including patching, MFA, passkeys, and strong EDR coverage.
Jun 29, 2026
ReliaQuest researchers reported that a credential-stealing malware campaign dubbed **DeepLoad** has been compromising enterprise IT environments through **ClickFix/QuickFix-style social engineering**. Attackers used fake browser prompts and error pages to trick users into running a malicious command, which launched a loader engineered to bypass multiple security tools. Researchers said the malware used heavily padded code and execution tied to a Windows lock-screen process, indicating **AI-assisted obfuscation at multiple stages** to reduce the effectiveness of static, file-based detection. Once installed, DeepLoad enabled **real-time keylogging**, stole credentials, spread to connected USB drives, and maintained hidden persistence that allowed the malware to re-execute days after apparent cleanup. The campaign's ability to survive standard remediation has raised concern that AI-generated variation is making malware more adaptable and harder to detect. Researchers urged defenders to emphasize **behavioral and runtime detection** over traditional static analysis to identify similar threats earlier.
Jun 29, 2026
Researchers and security teams reported mixed results on using large language models for offensive cybersecurity tasks, finding that current systems remain unreliable for autonomous exploitation while still proving useful in vulnerability research. Google Project Zero's **Project Naptime** evaluated whether LLMs could meaningfully support offensive security work, while the **CyberSecEval 2** benchmark measured model behavior across exploit generation, prompt injection, and code interpreter abuse. The benchmark found prompt-injection defenses remain weak, with successful attack rates of **26% to 41%** across tested models, and highlighted a safety-utility tradeoff in which stronger refusals can also block benign requests. At the same time, multi-agent LLM-assisted tooling demonstrated practical impact in software auditing. DARKNAVY said its **Argusee** architecture, which separates manager, auditor, and checker roles to reduce false positives and false negatives, uncovered **15 previously unknown flaws** in open-source projects and identified **CVE-2025-37891**, a high-severity Linux kernel USB vulnerability. The flaw affects the host side of the Linux USB stack from Linux `6.5` onward, can be triggered by a simulated USB device advertising the **MIDI2** protocol, and stems from improper length checks during MIDI1-to-UMP conversion that can cause an arbitrary kernel heap overflow; DARKNAVY reported reliable privilege escalation to root on Arch Linux before the issue was fixed.
Aug 6, 2026