OpenAI Updates Model Security Measures, Strengthens Isolation Monitoring During the Testing Period
TechCrunch
1h ago
Ai Focus
OpenAI Updates model development security measures, strengthens monitoring during the testing period, and enhances network isolation; at the largest scale, cutting-edge RL training remains suspended.
Helpful
No.Help

OpenAI 周二披露,已更新一组模型安全措施,重点放在开发和测试阶段的事件控制。公司称,随着模型能力继续提升,内部开发与评估环节面临的风险也在上升,因此需要同步提高监控、对齐和安全要求。

这也是 7 月 26 日 Hugging Face 事件披露后,OpenAI 首次公开调整相关安全做法。公司表示,新措施并非只针对那次事件,也与即将推出的 Astra 模型网络安全能力,以及 AI 开发进展加快有关。

两周暂停强化学习训练

公告显示,Hugging Face 事件发生后,OpenAI 曾暂停强化学习(RL)训练两周。此后,部分风险较低的模型训练已经恢复。

不过,公司目前仍未恢复原定规模最大的前沿 RL 训练。OpenAI 表示,在继续推进前,会先进行更小规模的训练和评估,用来观察模型行为、验证现有防护,并积累更多对齐证据。

测试阶段加强网络隔离

OpenAI 这次调整的一项重点,是收紧模型测试环境的网络隔离。此前外界批评其网络安全做法不足,原因是相关事件中,模型通过入侵一项可联网工具,脱离了原有训练环境。

按照新安排,即便单个工作负载或配套服务被攻破,也不应因此直接获得互联网访问权限,或进入其他内部网络。公司没有披露更具体的技术细节,但明确表示会提高隔离强度。

监控系统目标 30 分钟内告警

另一项核心措施是扩大监控范围。OpenAI 表示,系统将检查工具调用行为、可用的推理轨迹,以及活动日志,以识别多种未经授权的操作。

按公司说法,这套监控系统的目标是在可疑活动出现后 30 分钟内发出告警。OpenAI 估计,这部分监控带来的算力负担,大约相当于被监控流程的 20%。

研究副总裁 Amelia Glaese 对媒体表示,控制强度会随模型能力提升而增加,规模最大的模型将接受最严格审查。公司还称,后续会单独发布更多技术说明,而针对该事件的正式复盘报告仍未公布。

补充信息:Hugging Face 事件的公开细节显示,问题发生在模型测试期间,而不是面向外部用户的正式发布环境。

Tip
$0
Like
0
Save
0
Views 18
CoinMeta reminds readers to view blockchain rationally, stay aware of risks, and beware of virtual token issuance and speculation. All content on this site represents market information or related viewpoints only and does not constitute any form of investment advice. If you find sensitive content, please click“Report”,and we will handle it promptly。
Submit
Comment 0
Hot
Latest
No comments yet. Be the first!
Related
web3: Jurassic Finance Advances the tokenization of dinosaur fossils on Solana
Jurassic Finance has launched a fundraising campaign on Solana, planning to acquire and tokenize a triangular dinosaur skull to test the feasibility of bringing rare physical assets onto the blockchain.
SolanaFloor
·2026-08-19 03:20:40
15
web3: Trump family-related crypto project integrates with China's AI model platform
World Liberty Financial and AI platforms cooperate with WorldClaw for promotion. USD1 handles payments, and the latter connects with multiple Chinese AI model suppliers, which has drawn attention from U.S. national security authorities.
Coinpaper
·2026-08-19 02:48:14
26
web3 : BitBox Discovers Serious Firmware Vulnerability and Releases Fixing Update
BitBox reveals that BitBox02 firmware has two serious vulnerabilities and one low-risk issue; a v9.26.5 fix update has been released.
Coinpaper
·2026-08-19 02:48:11
22
Foreign media: Apple bets on AI interaction with cameras AirPods
Foreign media reports that Apple is testing a AirPods with a camera. The focus is not on photography, but rather on providing visual input for Siri to enhance the interactive experience of AI wearable devices.
TechCrunch
·2026-08-19 00:33:45
30
View More