Security researchers have found that a locally-run, uncensored AI model was able to modify a known credential dumping tool to evade detection by two commercial Endpoint Detection and Response (EDR) systems in a lab. The experiment, carried out by Project Black’s Eddie Zhang, shows how rapidly offensive tooling could be customized using generative AI without ever resorting to cloud APIs or prohibited prompts. This underscores a growing risk to endpoint security defenses.
What the experiment did
The offensive goal was to extract credentials from the Local Security Authority Subsystem Service (LSASS), a Windows process that stores authentication tokens in memory—a technique cataloged by MITRE ATT&CK as T1003.001. LSASS dumping is a dangerous threat vector because, once an attacker gains elevated system privileges, those in-memory tokens can allow lateral movement across targeted networks. The question: could an AI build an LSASS dumper that evades modern EDRs with minimal human direction? The answer: yes. But only under certain conditions.
Initial tests used closed-weight, guardrailed models such as Claude Opus 5, Opus 4.8, and Sonnet 5. All refused to generate a working dumper, even though the researcher had legit approval through a verification program. Only when switching to open models did results begin to emerge. With DeepSeek-V4-Flash 0731—an open-weight model—the researcher generated a working executable that could take a process ID, clone the target process in suspended mode via reflection, dump its memory in-memory, XOR-encrypt the dump, and write it to disk.
Evading EDR layers
Though the first build using DeepSeek tripped the EDR alarms, that changed after refinement. Eddie Zhang then used a locally-run, community-modified Qwen 3.8 model on powerful hardware. When prompted to increase stealth, the reworked binary passed unnoticed through two leading EDR systems in the lab. The changes: reduced access masks to the LSASS process, randomized delays during dumping, newer and non-standard output paths and filenames, modified process spawning, and removal of embedded strings in the binary. These changes targeted common heuristics EDR tools use—like detecting requests for high-privilege handles to LSASS, recognizing dump-writing behavior, or spotting process execution trees linked to credential dumping.
It’s not a universal bypass. No EDR vendors were named, and only two commercial products were tested under lab conditions. Configuration details weren’t shared, making it unclear how broadly applicable the evasion techniques are. That said, the attack demonstrates how accessible, community-built local AI models can be used to fine-tune dangerous code without the oversight or constraints typically found in hosted AI systems.
Mitigations and best practices
Defenses against LSASS dumping already exist—but this research shows they must be layered. Microsoft has several recommended hardening steps: enabling surface reduction rules focused on credential-stealing, protecting LSASS by running it as a Protected Process Light, activating Credential Guard, disabling outdated credential storage like WDigest, and limiting remote admin tools like RDP. Administrators should also manage local privilege escalation risk by limiting local admin rights, segmenting privileged accounts, using strong credential policies, and monitoring for anomalous handle requests or change in process behavior around LSASS.
Despite these options, however, incident responders need to recognize that EDR is not foolproof. Tools that tweak binary artifacts—filenames, entropy of process behavior, timing—can undercut much of what signature- or heuristic-based detection systems rely on.
The key takeaway: locally-deployed, uncensored AI models are starting to fill in gaps in attacker skill sets, enabling even less-skilled operators to produce customized, environment-specific offensive tools. Defenders must assume that some actors have access to such tooling. The arms race is accelerating.