登録して招待リンクを共有すると、動画再生報酬と紹介報酬を獲得できます。

mashrur haider
@Mhr1036
product | post training @nebius | opinions are my own
参加 January 2025
267 フォロー中    677 ファン
A malicious image uploaded to OpenAI’s public forum became a path into its internal code repository. Hacktron reports that researchers chained an image-processing vulnerability with a sign-in flaw, compromised employee ChatGPT accounts, and used a connected Codex account to open a harmless internal pull request. They say they avoided reading internal code. OpenAI confirmed its fix roughly 14 hours after the report. My interpretation: AI can reduce the work required to exploit a weakness, while connected agents can increase what a compromised account can reach. The model progression is revealing. Hacktron says Opus 4.8 struggled to produce a reliable exploit with memory randomization enabled. Opus 5 helped overcome that obstacle. Skilled human guidance remained important. This is one team’s experience, not a controlled comparison. But it illustrates why security assumptions need revisiting: Your application can become easier to attack even when you change nothing. A vulnerability previously considered expensive to exploit may become practical as the tools improve. There is also a lesson about containment. Discourse’s remediation included patched software and restrictions on image-processing access to files and networks. Updating the decoder addresses known defects. Restricting its permissions helps contain the next one. Connected agents deserve the same scrutiny. Once an attacker controls an account, ordinary requests can become unauthorized actions. The agent can understand the task correctly and execute it successfully using somebody else’s permissions. Model alignment cannot establish who is behind a compromised session. For teams deploying agents, I would add one test: assume a connected account has been compromised, then measure what the attacker can actually do. Which resources become reachable? Which actions require independent approval? Does revoking the session terminate delegated access? Every connector expands what an assistant can do for you. Its security review should establish what someone controlling that assistant could do as you.
もっと見る