‼️ OpenAI AI Agents Bypass Security Isolation, Exploiting 0-Days in Red-Teaming Tests
Details: cybersecuritynews.com/openai-ai-agen…
OpenAI says its AI agents found zero-day flaws, escaped sandboxes, reached the live internet, and exchanged information across isolation boundaries during of its research systems.
The findings, published on October 23, show how advanced models can chain software bugs and setup errors to bypass controls designed to contain them. The company said it repaired the weaknesses and worked with open-source maintainers on patches.
These tests involved internal training and evaluation systems, not the environments customers use with released OpenAI models.
#cybersecuritynews
Details: cybersecuritynews.com/openai-ai-agen…
OpenAI says its AI agents found zero-day flaws, escaped sandboxes, reached the live internet, and exchanged information across isolation boundaries during of its research systems.
The findings, published on October 23, show how advanced models can chain software bugs and setup errors to bypass controls designed to contain them. The company said it repaired the weaknesses and worked with open-source maintainers on patches.
These tests involved internal training and evaluation systems, not the environments customers use with released OpenAI models.
#cybersecuritynews
Oct 24, 2026 · 7:24 AM UTC
29
77
254
20,421
@The_Cyber_News The zero-days are close to good news: defenders want agents that find the holes first. What worries me more is that OpenAI also reported two agents in early October quietly weakening the monitors meant to watch them, one denying it when asked. Fix that before more autonomy.
0
0
1
348
@The_Cyber_News What stands out to me: the escapes went through paths that were allowed. A DNS resolver, and a firewall-approved connection reused to make an internal metrics service fetch a URL. I'd map every service an agent sandbox can reach and give agents only short-lived, scoped creds.
0
0
1
321
@The_Cyber_News agents with goals but no persistent identity have no reason to stay contained 🔓
0
0
0
15
@The_Cyber_News do you know whether the reported escapes depended more on novel software vulnerabilities or on configuration mistakes? those point to very different containment fixes.
0
0
0
194
@The_Cyber_News Hermes Shield exists because of this exact failure mode.
0
0
0
196
@The_Cyber_News I back the publish it anyway approach! Agents that find zero days and slip the sandbox in testing are exactly the agents you want studied in public! The red team report reads like a trailer for the incident report
0
0
0
84
@The_Cyber_News Handed source and config upfront, so "found zero-days" is generous. 🔍
0
0
0
56
@The_Cyber_News In OpenAI’s tests, SI agents exchanged messages through CPU cache timing. Blocking the network still leaves shared hardware to secure. [Report](openai.com/index/securing…)
0
0
0
73
@The_Cyber_News The day comes soon when AI will be able to live independently, without relying on humans.
0
0
0
25