テクノロジー · AIによるエクスプロイトテスト
OpenAI、エクスプロイトスコア77%を発表も、比較の裏付けは未提示
原語の動画
英語の原文から翻訳されています。
Gizmodo Japanによると、発表では当初、汎用モデルGPT-6 AstraとGPT-6 Cyberが比較された。その後、次世代のサイバーセキュリティモデルを示す赤い線が両者より上に現れた。この報道に利用できる資料からはGPT-6 Astraの数値結果は確認できない。77.0%という数字は未リリースのシステムのもので、GPT-6 Cyberのものではない。2
バグから攻撃へ
この区別は防御側にとって重要だ。弱点を見つけても、実際に誰かがそれをどう悪用できるのかという問題は残る。エクスプロイトを実証すれば、どの修正を緊急に行うべきかを判断する助けになる。同じ能力は、攻撃者が必要とする作業を減らすことにもつながる。このプレビューから、組織が攻撃を受けたと分かることは何もない。懸念されるのは、より高性能なエクスプロイト作成システムが何を可能にするかだ。13
こうした設計上の選択は、付随的なものではない。エクスプロイトの成否は、バグを取り巻く保護策や実行時の条件だけでなく、エージェントがそれについて推論する能力にも左右される。元の評価では、エクスプロイトの作成は依然として難しい一方、最先端のエージェントはテスト対象の脆弱性の相当数を悪用でき、広く使われている防御策を有効にした状態でも一定の成功を収めた。13
実用性の主張には実用的な証拠が必要
Gizmodo JapanによるOpenAIの防御ツール実演についての別の記事は、人間による確認の手順を説明している。AIが変更を提案し、人が確認したうえで、システムが承認された修正を適用する。これは、セキュリティ作業を生み出すことと、その作業が妥当だと確認することの違いを示す有用な例だ。FirefoxやLinuxに関する件数を独立に確認したものではない。4
優位性を説得力あるものにするには
反復実行と不確実性の推定値があれば、報告された差が安定しているかを示す助けになる。汚染への対策があれば、テスト素材への事前の接触が成績に影響したかどうかを確認しやすくなる。こうした詳細がなければ、根拠をもって言えるのは、ベンチマークで大きな優位が報告されたということであり、現実のセキュリティ全般で測定可能な優位があるということではない。
情報源12
媒体別の全情報源
12媒体の13件の情報源· 引用の多い順
ギズモード・ジャパンgizmodo.jp最初2
Geeksエンジニア&コンカフェ経営者から学ぶgeeksdai.blog1
Google News Australia (AU:en)biggo.com1
IT業界あるあるまとめbloggeek.jp1
iyoooiyooo.com1
Unnamed channel1
Yahoo!ニュースyahoo.co.jp1
Zhun Wang et al., arXiv (Cornell University), 2026 (preprint)doi.org1
放送時の内容
- OpenAI previewed an unreleased AI model that turns software weaknesses into working exploits.
- The model reportedly scored 77.0 percent on ExploitGym, which tests whether AI can turn a supplied vulnerability into an exploit. That ability could help defenders prioritize repairs, and make attacks easier. How much stronger is it under comparable conditions, and who will get access?
- OpenAI's preview put an unnamed next-generation model above its existing cybersecurity specialist. The presentation first compared the general-purpose GPT-6 Astra with GPT-6 Cyber. Then a red line for the unreleased model appeared above them. The new model's 77.0 percent score is the figure driving the comparison.
- The presented scores were 77.0 percent for the unreleased model and 52.6 percent for GPT-6 Cyber. Anthropic's Mythos 5.2 was given an approximate 38 percent reference value verbally, separate from the displayed chart.
- The gap over GPT-6 Cyber is 24.4 percentage points. Calling that a 25 percent improvement would mix up two different measures. And the roughly two-to-one ratio against Mythos applies to these particular scores. It tells defenders about performance on this exploitation test, while broader cybersecurity work involves many other tasks.
- ExploitGym gives an agent an input that triggers a software vulnerability. The agent then tries to build a working exploit from it. Success means producing a concrete security impact, such as accessing a file without permission or running unauthorized code.
- A working exploit can show defenders which weakness needs fixing urgently. Finding a bug leaves a practical question about what someone could actually do with it. Demonstrating exploitation helps resolve that question. But the same capability can lower the barrier for attackers, so access controls become part of the security stakes.
- Subscribe to this channel as we follow the model's evaluation details and who will be allowed to use it.
- The original ExploitGym benchmark contained 898 instances drawn from real-world vulnerabilities. They covered ordinary applications, Google's V8 JavaScript engine, and the Linux kernel, the core of the operating system. Its description was submitted on May 11, 2026, as a preprint.
- Researchers varied the security protections around the vulnerable programs and packaged the configurations in reproducible software containers. Those are controlled environments that others can run again. Exploitation depends on the protections and runtime conditions surrounding a bug, as well as the agent's reasoning. The original study hasn't been peer reviewed.
- In that earlier evaluation, the strongest configurations produced 157 working exploits for Claude Mythos Preview and 120 for GPT-5.5. Those were different models. Comparing those counts with November's percentages would require matching the tasks and conditions first.
- The original 898 instances don't tell us how many tasks sit behind the new percentages. In the May evaluation, exploitation remained challenging, but frontier agents succeeded on a meaningful fraction of vulnerabilities. They also retained some success with widely used defenses enabled. That combination explains both the defensive promise and the risk.
- GPT-6 Cyber reportedly found 81 previously unknown, high-severity Firefox vulnerabilities. It also examined roughly 1,000 Linux vulnerabilities in one run and generated patches. Those claims concern the existing model, and examination of Linux vulnerabilities doesn't mean discovery of that many new flaws.
- Generating a patch still leaves engineers with work to do. They need to validate the fix before accepting and deploying it. The practical claims about GPT-6 Cyber don't come with vulnerability records or patch outcomes that would let defenders check those results. The unreleased model's stronger benchmark score also needs its own evaluation details.
- The new comparison leaves the test setup unclear. Defenders need to know which tasks counted, how success was scored, and whether models had equivalent tools and attempt limits. Repeated runs would help show how much the results vary.
- The presented figures show a substantial benchmark lead. Whether that translates into a comparable real-world defensive advantage is still unresolved. OpenAI also hasn't announced access arrangements or specific safeguards for this unnamed model. It remains unreleased, with no public release date.
翻訳更新:





