Investing.comCodex · 实盘不做把握 3/10
北京时间·🕒 发布 20:32:43时区推断·⟳ 抓取 20:33:17·⚡ 决策 08-04 22:25:02 +1h51m·84.2 天前
Innodata发布AI安全编码训练数据套件
公司发布含12套数据集和评测系统的AI安全编码训练套件，但未披露订单、收入分成、客户采购或可量化商业化增量。
📄 原文Investing.com↗ 原文
Innodata releases AI cyber training suite for secure coding
BofA says this memory stock could double. Here are 5 reasons why NEW YORK - Innodata Inc. (NASDAQ:INOD) released the fi…
不做 Innodata（INOD.US）
仅为未披露金额和客户承诺的产品发布/营销型催化，无法量化短周期收入或利润增量，不做。
展开详情(成交明细 / 研究档案 / 原文全文 / 决策记录链)
研究档案
🔎 新鲜度:最早出现 2026-08-04 · 扩散 low · price-in 风险 medium ↗
关键理由
- 公司发布含12套数据集和评测系统的AI安全编码训练套件，但未披露订单、收入分成、客户采购或可量化商业化增量。
主要风险
- 标题与正文不一致，正文实际是公司产品新闻而非BofA记忆股研报，且产品发布类软催化短线可交易性弱。
原文全文
展开原文 · 2,078 字
BofA says this memory stock could double. Here are 5 reasons why
NEW YORK - Innodata Inc. (NASDAQ:INOD) released the first stage of its AI Cyber Training Suite on Monday, consisting of twelve datasets and evaluation systems designed to train AI coding agents to write and repair secure code, according to a press release statement.
The suite addresses security vulnerabilities in AI-generated code by training models to avoid known security flaws and patch existing vulnerabilities. It is available to model builders and enterprises starting today.
The datasets are built from thousands of real security flaws discovered over the past ten years across programming languages including Python, TypeScript, JavaScript, Rust, C, and Go. The flaws span platforms such as Linux, macOS, Android, Windows, AWS, and GCP. Innodata’s cybersecurity engineers reconstructed each flaw with a targeted attack inside a sealed offline copy of the vulnerable software.
The suite measures an AI’s ability to patch vulnerabilities by verifying that the original attack can no longer succeed against the fixed code while confirming the software still functions as intended.
Innodata tested ten AI models and agents on their ability to repair verified security flaws. The best open-weight AI models repaired at most 23% of flaws when shown exactly where each vulnerability was located. After one round of fine-tuning on a portion of the suite’s data, the model’s ability to patch vulnerabilities without guidance increased from 18.4% to 41.2%.
"For model builders, these evaluations and data sets enable coding agents - the models, harnesses, and tools - to ensure the code they generate is not just effective, but secure," said Yonatan Brander, who leads the cybersecurity data engineering team at Innodata.
The company states the suite enables enterprises to use AI to modernize legacy systems and patch identified vulnerabilities while minimizing the risk of introducing security flaws.
This article was generated with the support of AI and reviewed by an editor. For more information see our T&C.
NEW YORK - Innodata Inc. (NASDAQ:INOD) released the first stage of its AI Cyber Training Suite on Monday, consisting of twelve datasets and evaluation systems designed to train AI coding agents to write and repair secure code, according to a press release statement.
The suite addresses security vulnerabilities in AI-generated code by training models to avoid known security flaws and patch existing vulnerabilities. It is available to model builders and enterprises starting today.
The datasets are built from thousands of real security flaws discovered over the past ten years across programming languages including Python, TypeScript, JavaScript, Rust, C, and Go. The flaws span platforms such as Linux, macOS, Android, Windows, AWS, and GCP. Innodata’s cybersecurity engineers reconstructed each flaw with a targeted attack inside a sealed offline copy of the vulnerable software.
The suite measures an AI’s ability to patch vulnerabilities by verifying that the original attack can no longer succeed against the fixed code while confirming the software still functions as intended.
Innodata tested ten AI models and agents on their ability to repair verified security flaws. The best open-weight AI models repaired at most 23% of flaws when shown exactly where each vulnerability was located. After one round of fine-tuning on a portion of the suite’s data, the model’s ability to patch vulnerabilities without guidance increased from 18.4% to 41.2%.
"For model builders, these evaluations and data sets enable coding agents - the models, harnesses, and tools - to ensure the code they generate is not just effective, but secure," said Yonatan Brander, who leads the cybersecurity data engineering team at Innodata.
The company states the suite enables enterprises to use AI to modernize legacy systems and patch identified vulnerabilities while minimizing the risk of introducing security flaws.
This article was generated with the support of AI and reviewed by an editor. For more information see our T&C.
决策记录链 · 只追加
每次领取产生一条记录,重新决策追加一条并以 supersedes 指向被替代的记录(ADR 0014)。
这条决策早于决策账本(2026-10-08 起逐条追加),没有追加记录。