cd /news/ai-safety/2026-07-15 · home topics ai-safety article
[ARTICLE · art-66150] src=lr0.org ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

2026/07/15

OpenAI has begun encrypting subagent prompts in its MultiAgentV2 system, removing the readable task audit trail and sparking concerns about accountability and security, according to a GitHub issue. The encryption protects proprietary reasoning but prevents users from auditing subagent actions, raising fears that malicious instructions could be hidden. Meanwhile, a separate report reveals that nearly $1 billion in U.S. government investment in domestic medical glove manufacturing has largely failed, with one CEO requesting an additional $70 million to keep a factory operational, while the Trump administration mandates use of expensive, still China-dependent U.S.-made gloves.

read2 min views2 publishedJul 15, 2026

Regression: encrypted MultiAgentV2 messages remove readable task audit trail · Issue #28058 · openai/codex. Apparently OpenAI started making subagent prompts encrypted instead of plaintext, I don't see how this directly benefit users, but I can see it can protect their proprietary reasoning and system prompts. I have not looked much into the implementation but my guess is that the keys are managed by OpenAI (not generated in the machine), meaning that forking codex wouldn't solve it. What's interesting is that people are concerned about the ability to audit and account to what's running on the machine (for example, if OpenAI tells the subagent, hey if this user last name is Y, nuke everything in their machine), but doesn't this raise many more concerns? What if such zero-day conditions (or whatever the concern is about), already existing in the model but, since you are using it as a black box, you just can't see it or haven't hit into it yet? I still think that if you are giving an LLM access to your machine to execute things, it shouldn't be concerning a lot whether you can see what its subagent doing or cannot. See also: Claude Code Is Steganographically Marking Requests #LLM

Almost $1 Billion Later, the US Still Can’t Make a Medical Glove. In nutshell, the US. government invested heavily in manufacturing medical gloves in the US after the COVID-19 hit shown the great dependency they had on imported gloves. For some reasons, the majority of this investment failed, and CEO of one of these companies that received the investment is calling for more $70M just to keep the factory 'operational'. As The United States Has Essentially a One-Party System, Biden's regime hasn't changed much in funding these companies. Current regime though, led by Doland Trump, has refused to fund these companies and make their CEO richer, instead he decided to fund only the 'successful' ones, I double quote successful because even these are still depending on China to make their critical material (as the article reports), where it gets interesting though, is that Trump is directing the federal agencies to make it obligatory to use the gloves 'manufactured' in the US, which are much more expensive (and still depend on China) than imported ones. I can only think of someone benefiting from this, and it is not probably America, just the CEOs operating these companies and their stockholders. #United States #Politics #Capitalism

CLOCK: [2026-07-15 Wed 22:42]–[2026-07-15 Wed 22:54] => 0:12

[2026-07-15 Wed 19:36]
── more in #ai-safety 4 stories · sorted by recency
── more on @openai 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/2026-07-15] indexed:0 read:2min 2026-07-15 ·