Open-weight DeepSeek beats Google's gated Gemini model at finding bugs

DeepSeek V4 Pro, an open-weight model anyone can download, rediscovered 28 of 32 freshly disclosed CVEs for an 87.5% pooled vulnerability recall, topping every public closed model in a ten-model test run by cybersecurity firm Aikido, ahead of Anthropic's Opus 5 and Alibaba's Qwen 3.8 at 81.3% each. Three DeepSeek Pro runs cost Aikido roughly $295 total, while a single pass from Opus 5, Grok 4.6 or OpenAI's Sol ran $450 to $590 and found fewer bugs. The result undercuts Google's rationale for routing Gemini 4 Argon exclusively to vetted defenders through its Fairwind Program over cyberattack fears, especially since Google made gemini-4-argon generally available on October 13, about two weeks after the restricted rollout began.

Google kept its most capable model behind a vetted-access program over cyberattack fears. A free-to-download rival is already outscoring every closed model at the exact task Google worried about. On September 30, Google announced Gemini 4 Argon, its most capable model yet, and then did something unusual: it didn't let the public use it. Instead, Google routed the model exclusively to a small group of vetted cybersecurity defenders through what it calls the Fairwind Program, while it worked through the U.S. government's voluntary pre-release review process. Koray Kavukcuoglu, Google DeepMind's chief AI architect, wrote that "safely releasing frontier capabilities at this level requires a phased approach." The company's stated fear, reported by multiple outlets including The Star and Ary News, was that Argon's skill at finding software vulnerabilities could just as easily help someone break into a hospital, a bank, or a government system as help someone defend one. That caution looks a lot less airtight than it did three weeks ago. According to Aikido, a cybersecurity firm that ran ten AI models against 32 freshly disclosed CVEs to test genuine vulnerability reasoning rather than memorized answers, DeepSeek V4 Pro, an open-weight model anyone can download and run themselves, topped every public closed model on pooled vulnerability recall. Pooling three runs, it rediscovered 28 of the 32 vulnerabilities, a hit rate of 87.5%, ahead of Anthropic's Opus 5 and Alibaba's Qwen 3.8, both of which landed at 81.3%. And it did it cheap: three DeepSeek Pro runs cost Aikido roughly $295 total, while a single pass from Opus 5, Grok 4.6 or OpenAI's Sol ran $450 to $590 and still found fewer bugs. Here's the part that undercuts Google's own logic. The precise capability Google cited as too risky for open release, finding exploitable flaws in real code, is the same capability an open-weight competitor is already demonstrating in public, with a published methodology and a named firm standing behind the numbers. Nobody at Google can un-ring that bell by keeping Argon in a velvet-rope program. It's worth being precise about what "too dangerous to release" meant in practice, because it wasn't a permanent lockdown. Argon reached general availability on October 13, just two weeks after the restricted rollout began. Google made gemini-4-argon available on the paid tier of its API, rolled it out to Google AI Pro and Ultra subscribers in the Gemini app and Search, and brought it into Google Antigravity for Ultra subscribers. So the actual gap between "cyber defenders only" and "anyone with a credit card" was roughly two weeks, not an indefinite hold. DeepSeek Ships V4 Pro, Rivaling Claude and GPT-5 for a Fraction of the Cost https://startupfortune.com/deepseek-ships-v4-pro-rivaling-claude-and-gpt-5-for-a-fraction-of-the-cost/ DeepSeek pushed its V4 Pro model to general availability with no announcement, and it now scores 80.6% on SWE-bench Verified and posts the highest Codeforces rating of any model tested. It does all of that while charging roughly 46 times less than Claude Opus 4.7 on a blended token rate. - deepseek v4 pro pricing comparison https://startupfortune.com/deepseek-ships-v4-pro-rivaling-claude-and-gpt-5-for-a-fraction-of-the-cost/ - open weights coding model alternative https://startupfortune.com/deepseek-ships-v4-pro-rivaling-claude-and-gpt-5-for-a-fraction-of-the-cost/ On the benchmarks Google did disclose, Argon leads or ties in 13 of 18 categories, scoring 68.0% on CWE-bench v1 for vulnerability remediation and 85.8% on source-code vulnerability scanning across 20 languages, according to Google's own released comparison data. It trails GPT-6 Astra by 10.5 points on two coding-focused benchmarks and trails Claude Opus 5.5 by 9 points on a separate agentic test. None of that is nothing. Argon is a genuinely strong model. But strong and uniquely dangerous are different claims, and the two-week gating window suggests Google itself didn't treat the risk as severe enough to hold the line past a government review cycle. Meanwhile the open-weight field hasn't been standing still waiting for permission. DeepSeek isn't alone: Qwen, Kimi, and Zhipu's GLM-5.3 all showed up in Aikido's results with strong, consistent vulnerability-finding performance, and in a separate Cyber Defense Benchmark cited by multiple outlets, Kimi K3 hit 33% coverage and DeepSeek 4 Flash hit 29.7% at a median $0.17 per investigation, trailing only the open-weight leader GLM 5.2. These aren't fringe experiments. They're models security teams can already run today, on their own hardware, with no vendor gate at all. Also read: SoftBank's Son Plans $50 Billion Malaysia Tech Push, PM Says https://startupfortune.com/softbanks-son-plans-50-billion-malaysia-tech-push-pm-says/ • DeepSeek's cheap new model is making GPU scarcity worse, not better https://startupfortune.com/deepseeks-cheap-new-model-is-making-gpu-scarcity-worse-not-better/ • Cerebras bets on wafer-scale chips as Nvidia's GPUs get scarce and pricey https://startupfortune.com/cerebras-bets-on-wafer-scale-chips-as-nvidias-gpus-get-scarce-and-pricey/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.