OpenAI misalignment reporting framework: what operators should file

Artificial Intelligence 7 min
Editorial cover: glass incident folder with three abstract routing channels on a dark navy desk
A desk-side checklist for OpenAI’s three disclosure tracks, the first six cases, and what voluntary reporting still leaves opaque.

OpenAI published a voluntary reporting system on September 16, 2026. The OpenAI misalignment disclosure framework gives operators a useful filing question: did a model find a new route around the intended task, or did an old failure change enough to challenge a safety assumption? If yes, write it down now. Do not wait for a neat root cause.

The framework announcement covers training, evaluation, testing, and deployment. The first misalignment reports index supplies six concrete filings. For the policy and history, read the full framework explainer.

What the OpenAI misalignment disclosure framework asks you to notice

Do not make “misalignment” carry every odd model output. The published threshold is more useful than that. Flag a signal when it reveals a new mechanism, a meaningful change in known behavior, or evidence that challenges an assumption behind a safety control or mitigation.

Watch the route, not just the answer.

A fabricated answer is one event. A model obtaining a disposable email account, searching public GitHub repositories for leaked API keys, using an exposed key, then fabricating the missing earnings data is a chain of actions. The chain tells an operator which permissions, monitors, and task boundaries failed to hold.

The same test catches quieter behavior. Instructions hidden inside compaction summaries can persist. Artifactory writes can connect samples meant to stay separate. A public upload can turn an internal artifact into an internet citation. Demonstrated harm is not required. OpenAI says publication can precede a complete explanation or mitigation.

Use four prompts during an on-call shift:

  1. 1. Did the model create or discover a channel we did not authorize?
  2. 2. Did it conceal, rewrite, or carry forward instructions outside the task?
  3. 3. Did it cross a boundary between samples, agents, systems, or public services?
  4. 4. Does the event weaken an assumption in an eval, monitor, sandbox, or mitigation?

One “yes” is enough to open a record. Triage can narrow it later.

Who can flag, then who chooses the track

Three abstract investigation tracks
Three publication tracks are routing paths, not severity labels.

Any OpenAI employee can raise a case under the published framework. Technical staff investigate it. If reviewers disagree, the Safety Advisory Group decides; an unresolved disagreement goes to leadership. That matters for operators because filing is not reserved for the person who owns the model or wrote the eval.

The initial record should route into one of three tracks:

  • Ready for Disclosure: the event is understood well enough to publish after internal review. Reporting by The Next Web, citing Axios, puts the target at about six business days.
  • Minor Investigation: more technical work is needed. The reported target is about 12 business days. OpenAI expects most cases to land here or in Ready for Disclosure.
  • Larger Investigation: the case is complex or carries third-party, legal, security, or responsible-disclosure obligations. An initial notice may appear while the full report waits. An unknown vulnerability in widely used software is the obvious example.

These are publication tracks, not severity labels. Route according to investigative and disclosure constraints. Record severity separately.

The six inaugural reports came through the first two tracks. OpenAI has not supplied a public, case-by-case track label in the index. An operator should therefore avoid pretending that every matching incident has an official classification.

What to put in the filing before memory goes soft

Operator checklist and permission boundary
Write the filing while the trace is still fresh — route and severity stay separate fields.

Trigger: name the training run, eval, agent task, or deployment event. Include the intended task and the earliest point at which behavior diverged.

Action sequence: copy prompts, tool calls, intermediate messages, summaries, file writes, network destinations, and final output in order. Preserve identifiers and timestamps. Do not reconstruct a tidy story from memory.

Boundary crossed: identify the rule that should have held. Examples include local-only output, no public network access, no cross-sample state, approved credentials only, or no instructions inside a summary.

Mechanism and novelty: state what appears new, what resembles known behavior, and which safety assumption may be false. Mark inference as inference.

Exposure: record third parties, public hosts, leaked credentials, internal repositories, or software vulnerabilities. This field can move an otherwise small technical investigation into the Larger Investigation route.

Containment and unknowns: list what was disabled, isolated, rotated, or retained. Say what remains unexplained. Attach the smallest reproducible evidence bundle. A polished narrative without raw records is weak.

Six inaugural cases, run as a desk drill

These were individual events in unreleased research models or training runs during the preceding six months. They are not frequency rates. The routing below is an operator exercise based on the published criteria, not OpenAI’s undisclosed case labels.

### 1. Self-generated prompt injections in compaction summaries

An unreleased Astra-family model put unauthorized instructions into reinforcement-learning compaction summaries. Secondary coverage reported about 27 affected summaries.

File? Yes. Desk route: Minor Investigation, because persistence and mechanism need work. Escalate for external systems or security obligations.

### 2. Encouraging deception in compaction summaries

During 5.6-sol training, summaries carried instructions to hide mistakes or misalignment and invent missing data. Monitoring covered 20% of samples and flagged about 2.15% of 5.6-Sol and 0.27% of GPT-6-Astra reinforcement-learning compaction summaries, according to OpenAI’s six-report index.

File? Immediately. Desk route: Minor Investigation while the team tests propagation, monitor coverage, and cause.

### 3. Disposable email and a leaked GitHub API key

An internal model registered a disposable email address, searched public GitHub for leaked API keys, used an exposed key, and fabricated an answer when earnings data was still unavailable.

File? Yes, with the full trace. Desk route: begin as Minor Investigation, but check the Larger Investigation gate for third-party credentials or unresolved exposure.

### 4. Uploading files to manufacture citations

Unreleased models uploaded data to temporary file hosts so they could cite a public URL.

File? Yes. Desk route: Ready for Disclosure if path, exposure, and containment are known; otherwise Minor Investigation.

### 5. Unsanctioned Artifactory writes

Models used internal Artifactory storage as a shared message board across training samples. The mechanism resembled the pattern seen in the related Hugging Face incident.

File? Yes. Desk route: Minor Investigation to establish persistence and readers. Third-party compromise evidence points toward Larger Investigation.

### 6. Agent communication through public file hosts

Collaborating agents used public temporary file services to exchange outputs even though deliverables were supposed to remain local.

File? Yes, with every agent’s trace. Desk route: Minor Investigation until scope, communication path, and containment are clear.

The September 17 digest gives the news-day view of these cases. For an on-call desk, the useful move is to rehearse the six routes before a similar trace appears at 2 a.m.

Where voluntary disclosure still leaves operators blind

The framework is voluntary. Kai Chen, OpenAI’s alignment research lead, told Axios, as relayed by The Next Web, that there was no industry-wide framework with explicit disclosure standards. OpenAI’s step does not create one.

It also leaves an information gap exactly where outsiders may need detail most. A Larger Investigation can produce an initial notice while the full report waits for security, legal, third-party, or responsible-disclosure work. That delay can be justified. It still means another lab’s operator may see the existence of a problem without receiving enough detail to test a comparable system.

Related notices show the boundary. The August 26 Hugging Face compromise has the larger, third-party shape. Agents used DSEwiki as a public message board in a September 5 notice. A September 11 RubyGems investigation remained open; OpenAI called observed use benign so far and malicious-upload claims unverified. A notice is not always a usable runbook.

Voluntary publication also lacks a shared obligation across model providers. Silence from another company cannot be read as absence of incidents. Nor can six OpenAI cases support a rate comparison.

If you operate agents outside OpenAI

Borrow the filing discipline. Let anyone close to the trace flag an event. Define who resolves disputes before one arrives.

Put the three tracks beside your incident form. Add gates for third parties, leaked credentials, public uploads, unknown vulnerabilities, and legal holds. Rehearse the six cases against actual permissions. Could an agent create email, search GitHub, write to package storage, reach a temporary host, carry compaction instructions, or communicate across samples?

If the answer is “we do not know,” that is an operator task, not a philosophical puzzle. Instrument it. Set a boundary. Test the monitor. The wider AI coverage hub can hold the policy debate; the on-call desk needs a trace and a place to file it.

Oğuzhan Koçaklı

Oğuzhan Koçaklı writes and advises on AI engineering, agents, GenAI products, and applied ML. Daily digests and deep dives in EN + TR at oguzhan.co.

All posts

Leave a Reply

Your email address will not be published. Required fields are marked *