Two months after Kimi K3's release, no recorded harm from its open cyber capabilities, but open-weight safety gaps spark debate

About two months after Kimi K3's open-source release sparked community alarm over its built-in cyber offense capabilities, there is still no documented, attributable real-world harm from K3—providing an empirical reference point for the debate over whether releasing cyber attack capabilities in open frontier models inevitably leads to abuse.

Confirmed

• Cautious assessments from AISI/CAISI and others put K3's cyber capabilities at roughly Opus 4.6 level.
• AlexBarry4 argues in a long post that Mythos/5.6-Sol represents a huge leap in cyber capability; what currently prevents frontier models from being used for large-scale destructive attacks is mainly defenses like Anthropic/OpenAI's deployed classifiers, a layer open-source models lack.
• xeophon proposed an experiment: run SWE and cyber benchmarks on safety-stripped abliterated models, suspecting most merely "play the part," or that abliteration itself reduces capability; both sides agreed to revisit in a few months.

Not yet confirmed

• anthonyronning cited K3's two months without real-world harm to push back on concerns; xeophon countered that crypto has been hit by major hacks one after another, but the link between the two lacks evidence—the point remains disputed.
• Whether the situation materially changes if open-source models reach Mythos/5.6-Sol's frontier cyber attack levels within 1-2 months is unanswered; xeophon plans to check back in 5-6 months.
• There is disagreement over the GLM abliterated API's actual effectiveness: xeophon thinks the excitement is overblown, since prompt injection alone can make closed-source models bypass refusals.

Why it matters

• This case offers a rare empirical window into the open-source safety governance debate: releasing a model with high cyber capability open-source does not necessarily lead to immediate abuse, but the classifier layer of protection is indeed absent from the open-source ecosystem—the risks as open capabilities approach the frontier remain to be tested.

2026-09-10 ~ 2026-09-10 · 7 related posts

• [source] Two months after release, open-weights Kimi K3 shows no documented harm despite cyber capabilities — xeophon · 2026-09-10
• Two months after Kimi K3's release, no documented harm despite its open cyber capabilities — anthonyronning · 2026-09-10
• [source] Open model cyber capabilities near Opus 4.6 level: no K3 harm yet, but trend worries safety watchers — xeophon · 2026-09-10
• [source] Only Vendor Classifiers Hold Back Frontier AI Cyber Attacks — and Open Models Have None — AlexBarry4 · 2026-09-10
• If Open Models Hit Frontier Cyber Levels in Months, 'Check Back in 5-6 Months' — AI Cyber Attack Debate — xeophon · 2026-09-10
• Testing whether "abliterated" models are just larp: run them on SWE and cyber benchmarks — xeophon · 2026-09-10
• Devs suspect 'abliterated' model variants are mostly larp, propose SWE and Cyber benchmarks to prove it — AlexBarry4 · 2026-09-10