Open-Source Models Are Losing Security Capabilities: Concerns Over Kimi and Tencent's New Models

Are Open-Source Models Losing Their "Security" Capabilities?

Recently, a post on Reddit caught the attention of the security community. The poster pointed out that two newly released open-source models, Kimi K3.1 and Tencent's Hy4, have both had cybersecurity, reverse engineering, and vulnerability patching capabilities "stripped" from them. The poster believes this is becoming a troubling pattern among recent Chinese open-weight models.

If security data is stripped from local models, developers and security researchers will struggle to audit code vulnerabilities, reverse engineer binaries, or patch other vulnerabilities on their own hardware. This could push them toward closed-source APIs, where false-positive shutdown mechanisms are a poor experience for everyone.

Why Is This Worth Watching?

The value of open-source models in the security field lies precisely in their auditability and local deployment capabilities. Security teams need to be able to inspect a model's training data, inference logic, and output behavior to confirm there are no hidden backdoors or biases. When a model is deliberately weakened on security-related tasks, that foundation of trust is shaken.

The poster specifically mentioned the Hugging Face incident. In that case, it was open-source models that stepped up, not the closed-source "gatekeepers." This hints at a key point: when responding to security incidents, the open-source community's rapid response and transparency are often more reliable than commercial closed-source solutions. Weakening the security capabilities of local models ultimately hurts defenders.

Open-weight models refer to models whose parameters (weights) are publicly released and allowed to be downloaded and run locally, as opposed to closed-source models that only offer API access. Open weights are not the same as fully open source—training data, training code, and data processing pipelines are often not disclosed, so "auditability" actually varies by level: researchers can inspect weights, reproduce inference behavior, and conduct red-team testing, but may not be able to trace the composition of training data. This also means that when a model has certain capabilities "trimmed," it is difficult for outsiders to distinguish whether this is due to changes in training data ratios, safety alignment during post-training (such as refusing certain categories of requests), or deliberate removal of domain-specific corpora from the data pipeline. For security research, weights are visible but "black box" elements remain—this is precisely why the community remains sensitive to capability degradation.

Community Concerns and Divergent Views

The poster raised an open question: will more open-source labs go down this path? From the tone of the post, he is pessimistic about this and hopes other labs will not follow suit.

It should be noted that this observation currently comes mainly from a single community source and has not been officially confirmed. The "trimming" of model capabilities may be due to safety alignment considerations, or it may be caused by changes in training data or evaluation benchmarks. But regardless of the cause, the community's concern about the degradation of open-source model security capabilities is real.

Safety alignment typically refers to using supervised fine-tuning, RLHF, or rule-based refusals to make models behave cautiously or refuse when dealing with offensive cybersecurity operations (such as exploit writing or malware analysis assistance). The intent of such alignment is to prevent model misuse, but its boundaries are difficult to define precisely: defensive work such as vulnerability patching, reverse engineering, and malware forensics overlaps heavily with offensive operations technically, and coarse-grained refusal strategies can easily shut out legitimate security research needs as well. Therefore, "capabilities being stripped" and "safety alignment tightening" are often difficult to distinguish on the surface, and evaluation on specific tasks (such as CTF-style problems, patch generation, binary analysis) is needed to determine whether a model is actually missing capabilities or is simply triggering refusals.

Impact on Developers

For developers who rely on local models for security work, this trend means they need to more carefully evaluate how new models perform on security tasks. If mainstream open-source models continue to weaken related capabilities, security teams may be forced to choose between "using local models with limited capabilities" and "relying on closed-source APIs that may produce false positives."

The poster emphasized that this is "bad for everyone." Security defense should not become a casualty of competition over model capabilities. The open-source community needs to continue monitoring this trend and push model providers to remain transparent about security capabilities.

False-positive shutdown refers to closed-source API providers directly rejecting requests or banning accounts based on content moderation policies when they suspect a violation, even though the moderation judgment itself may be wrong. For security researchers, the problems caused by such mechanisms are not limited to a single failed request: account bans can interrupt ongoing analysis work, and appeal processes are usually opaque and slow. In contrast, locally deployed open-weight models do not have server-side moderation or remote shutdown issues, and researchers can fully control invocation methods and data flows—this is one of the important practical reasons why security teams prefer local models.

One-line Summary

The community alleges that Kimi K3.1 and Tencent Hy4 have removed security-related capabilities, raising concerns about the security value of open-source models.

Paragraph Summary

A Reddit community post pointed out that the newly released open-source models Kimi K3.1 and Tencent Hy4 appear to have removed capabilities related to cybersecurity, reverse engineering, and vulnerability patching, and the poster believes this is becoming a bad pattern among Chinese open-weight models. The article analyzes the reasonableness of this concern: the core value of open-source models lies in auditability and local deployment, and security teams rely on them to check for backdoors, audit vulnerabilities, and perform reverse analysis; if capabilities are weakened, researchers may be forced to turn to closed-source APIs and bear the risks brought by false-positive shutdowns. The article also notes that this observation currently comes only from a single community source and has not been officially confirmed, and that capability "trimming" may also stem from safety alignment or changes in training data. The core conclusion is: regardless of the cause, the degradation of open-source model security capabilities deserves continued attention, and providers should remain transparent.