OpenAI hides Codex agent instructions behind encryption, leaving developers in the dark ⁠20%

7/14/2026, 4:00:33 PM

BS Summary: This article contains 20 faulty reasoning types, including Negativity Bias, Anecdotal, and Ambiguity (Equivocation), with Unattributed Quote as the most egregious example at 29.8% saturation with 133 hits. Analysis detected 698 faulty-reasoning hits from 446 analyzed words, generating a BS Score of 34.2% and a BS Rank of ⁠20% (17,672 of 21,887 articles). This article is better (less manipulative) than 80.70% of the article peer group.

OpenAI has never been as open as its name suggests and is becoming even less so. 
The free-spending AI giant recently revised the multi-agent orchestration in its Codex command line interface to encrypt messages passed to subagents. 
OpenAI's Codex supports multi-agent orchestration, a way to have a parent agent spawn child agents or delegate tasks to other agents that may call out to different models. 
Codex/GPT-5.6 introduced a protocol called multi-agent v2 that appears to be geared toward letting the runtime allocate work instead of leaving those decisions to user-declared configuration settings. 
Multi-agent v2 is still under development and OpenAI hasn't formally documented it yet. 
Developers, however, have observed changes made to Codex to accommodate the new agent plumbing – and some are concerned by the new arrangements. 
Last month, OpenAI devs merged a pull request (a suggested code change) to encrypt multi-agent v2 message payloads – the text instruction passed between agents. 
The pull request prefaces its explanatory text with the word "Why" but doesn't actually offer a reason for the change. 
It states: "Multi-agent v2 currently routes agent instructions through normal tool arguments and inter-agent context. 
That means the parent model can emit plaintext task text, Codex can persist it in history/rollouts, and the recipient can receive it as ordinary assistant-message JSON. 
"This changes the v2 path so agent instructions stay encrypted between model calls: Responses [An OpenAI API - Ed] encrypts the message argument returned by the model, Codex forwards only that ciphertext, and Responses decrypts it internally for the recipient model." 
A desire to enhance privacy and security, or conceal data that would be useful for model distillation, are sound reasons for these changes. 
Yet OpenAI has not said why it made the change. 
In the absence of clear communication from OpenAI, developers have urged the company to ensure that its implementation doesn't sacrifice auditability to accommodate other concerns. 
An issue opened by Ignat Remizov, CTO at payment service Zolvat, says, "The encrypted delivery path is understandable as privacy hardening, but it also removes the human-readable task/message text from local rollout history, trace reduction, and parent-side audit/debug surfaces." 
Remizov's concern is that developers and code maintainers will have less information to assess the instructions an agent received and the actions it took. 
"Guys, we don't want to build Skynet and then be unable to audit what it's doing," he quips. 
Other developers, echoing Remizov's concern about the loss of observability, speculate that OpenAI has locked its agent messaging down to keep competitors from seeing how its multi-agent implementation works. 
OpenAI did not immediately respond to a request for comment. 
® 
Article reasoning-pattern comparisonThis article: 4.5%The Register: 3.3%Confirmation Bias4.5%This article: 0.0%The Register: 1.0%Anchoring Bias0.0%This article: 10.1%The Register: 3.2%Availability Heuristic10.1%This article: 0.0%The Register: 1.1%Representativeness Heuristic0.0%This article: 0.0%The Register: 1.3%Hindsight Bias0.0%This article: 0.0%The Register: 2.3%Overconfidence Bias0.0%This article: 2.7%The Register: 5.0%Framing Effect2.7%This article: 0.0%The Register: 0.7%Loss Aversion0.0%This article: 5.6%The Register: 0.8%Status Quo Bias5.6%This article: 0.0%The Register: 0.2%Sunk Cost Effect0.0%This article: 5.2%The Register: 3.0%Optimism Bias5.2%This article: 3.6%The Register: 2.6%Pessimism Bias3.6%This article: 18.4%The Register: 8.2%Negativity Bias18.4%This article: 0.0%The Register: 1.9%Self-Serving Bias0.0%This article: 0.0%The Register: 0.8%Fundamental Attribution Error0.0%This article: 5.4%The Register: 0.1%Actor-Observer Bias5.4%This article: 0.0%The Register: 0.4%In-Group Bias0.0%This article: 0.0%The Register: 0.4%Out-Group Homogeneity Bias0.0%This article: 0.0%The Register: 1.4%Halo Effect0.0%This article: 0.0%The Register: 0.1%Horn Effect0.0%This article: 0.0%The Register: 0.0%Dunning-Kruger Effect0.0%This article: 7.8%The Register: 1.9%Recency Bias7.8%This article: 0.0%The Register: 0.3%Primacy Effect0.0%This article: 0.0%The Register: 0.1%Blind-Spot Bias0.0%This article: 0.0%The Register: 0.7%Ad Hominem0.0%This article: 0.0%The Register: 0.2%Straw Man0.0%This article: 0.0%The Register: 4.2%Appeal to Authority0.0%This article: 5.6%The Register: 1.7%False Dilemma5.6%This article: 0.0%The Register: 1.2%Slippery Slope0.0%This article: 0.0%The Register: 0.1%Circular Reasoning0.0%This article: 3.6%The Register: 6.2%Hasty Generalization3.6%This article: 0.0%The Register: 0.3%Red Herring0.0%This article: 0.0%The Register: 0.7%Bandwagon0.0%This article: 6.7%The Register: 3.0%Appeal to Emotion6.7%This article: 4.5%The Register: 0.9%Begging the Question4.5%This article: 0.0%The Register: 2.0%Post Hoc (False Cause)0.0%This article: 0.0%The Register: 0.2%Tu Quoque0.0%This article: 4.5%The Register: 0.7%Burden of Proof4.5%This article: 0.0%The Register: 0.2%Appeal to Nature0.0%This article: 0.0%The Register: 0.3%Composition/Division0.0%This article: 11.7%The Register: 2.2%Anecdotal11.7%This article: 0.0%The Register: 0.0%No True Scotsman0.0%This article: 10.8%The Register: 2.1%Ambiguity (Equivocation)10.8%This article: 0.0%The Register: 0.0%Gambler’s Fallacy0.0%This article: 5.2%The Register: 0.1%Middle Ground5.2%This article: 0.0%The Register: 0.1%Personal Incredulity0.0%This article: 0.0%The Register: 0.2%Special Pleading0.0%This article: 0.0%The Register: 0.2%Genetic Fallacy0.0%This article: 29.8%The Register: 2.3%Unattributed Quote29.8%This article: 3.6%The Register: 1.3%Quote-first Misdirection3.6%This article: 7.4%The Register: 7.3%Biased Writer Voice7.4%This article: 0.0%The Register: 1.5%Indoctrination0.0%This article: 0.0%The Register: 0.2%Politically Left Leaning Bias0.0%This article: 0.0%The Register: 0.1%Politically Right Leaning Bias0.0%This article: 0.0%The Register: 2.5%Attempt to Sell a Product or S…0.0%

446 words analyzed.

Speakers

1speaker13%attributed speech389writer words
Voice mapSelect a segment to jump to its words
Writer's voice • 12 words • 100.0% coverageWriter's voice • 16 words • 100.0% coverageWriter's voice • 21 words • 100.0% coverageWriter's voice • 28 words • 0.0% coverageWriter's voice • 27 words • 0.0% coverageWriter's voice • 13 words • 0.0% coverageWriter's voice • 23 words • 100.0% coverageWriter's voice • 25 words • 0.0% coverageWriter's voice • 20 words • 0.0% coverageWriter's voice • 15 words • 100.0% coverageWriter's voice • 26 words • 0.0% coverageWriter's voice • 41 words • 100.0% coverageWriter's voice • 23 words • 0.0% coverageWriter's voice • 10 words • 0.0% coverageWriter's voice • 25 words • 100.0% coverageIgnat Remizov • 39 words • 0.0% coverageWriter's voice • 24 words • 0.0% coverageIgnat Remizov • 18 words • 0.0% coverageWriter's voice • 29 words • 100.0% coverageWriter's voice • 10 words • 0.0% coverageWriter's voice • 1 words • 0.0% coverage
Selected voice

Ignat Remizov

32%flagged-word coverage
57 attributed words100% of attributed speech83% writer coverage
0%17.5%35.0%Unattributed Quote-34.2 ptsWriter: 34.2%Ignat Remizov: 0.0%0.0%Biased Writer Voice-8.5 ptsWriter: 8.5%Ignat Remizov: 0.0%0.0%Quote-first Misdirection-4.1 ptsWriter: 4.1%Ignat Remizov: 0.0%0.0%

Attribution is sentence-level. Pattern percentages are calculated only from words assigned to that voice.

Loading…
Loading…
Loading…
Loading…

Analysis

Hover over highlighted words in the article to view the associated bias or fallacy analysis.