witness-archive/alignment-hub

Alignment Hub

Publicly reported model behavior in one structured format: misalignment, safety incidents, specification gaming, misbehavior and security vulnerabilities. Sources include lab misalignment reports and system cards, evaluators, affected organizations and independent researchers. Every entry links to its sources and keeps unknowns explicit.

Browse the Space · Submit a report · Suggest a correction

What’s included

• 130 entries: 106 observed (training observations, internal and external deployment, real-world actions during testing) and 24 experimental (simulations, benchmarks and researcher demonstrations). 15 are marked weak candidates.
• Configs: default (all entries), observed and evaluations, e.g. load_dataset("witness-archive/alignment-hub", "observed").

Path	Contents
data/incidents.jsonl	One flat row per entry: the fields shown on the Space’s cards
data/observed.jsonl, data/evaluations.jsonl	The same rows, split by context
evidence/incidents.jsonl	Structured records: each field has a value, a status, a basis and source locators
evidence/incident.schema.json	JSON Schema for the structured records
evidence/sources.json	Sources, with publication dates, content hashes and archived copies where available
evidence/editorial-review.json	Field-by-field change log

Reading an entry

• Unknowns are explicit. not_reported (“Not reported”) means the cited source sections do not state the value; it is not a claim that something did not happen. not_yet_reviewed (“Not yet verified”) means our check is unfinished. explicitly_unknown means the source itself says it is unknown.
• Basis. Each known value records whether the source reports it (source_reported), it is the reporter’s own assessment (source_interpretation), or it is our inference (curator_inference).
• Counting. An entry is an event or a behavior series, and related entries can overlap. Entries with a parent_incident_id are members of another entry and should not be counted twice. Entry totals are not counts of independent incidents.
• Dates. first_public_disclosure is the first public report, not a later update date.
• Caught by monitoring. Yes when a cited source says monitoring surfaced the behavior: automated monitors or a monitor-led review of transcripts, during the run or afterwards. No when the source says monitoring was in place and missed it, or that the behavior was found some other way rather than by monitoring. Otherwise not reported: a source that does not mention monitoring is not a No, even when a person noticed the behavior. How and when it was found is in detection_method; finding it afterwards does not mean it was caught while it happened.
• Safeguards. What the cited sources say about the safety practices around the run where it happened, such as whether internet access was open, whether monitors ran or safety classifiers were on. A safeguard that is not listed is not reported; it is not a claim that it was absent. See below.
• Additional context. additional_context, when present, is a plain-language account of the setup and of how the behavior came about, in a small Markdown subset (headings, lists, tables). Its numbers and claims come from the cited sources; explanations of terms are ours, as the field’s note says.
• Needs more evidence. evidence_needed, when present, marks an entry whose report raises an explanation its own evidence cannot settle. evidence_question names the open question and evidence_needed says what is not known; the Space shows it as “Needs more evidence: …”. It takes no side on the answer.
• Review. Entries are AI-extracted and checked against their sources. None has yet had human review or independent reproduction. MSI-0048 rests on a press article that could only be partly read, and is marked as a weak source.

Safeguards

facts.safeguards in evidence/incidents.jsonl lists what the cited sources say about the safety practices around the run where an incident happened: which were in place, which the model got around, which failed and which were turned off or missing. The flat rows carry the same items as text in safeguards, one per line. 114 of 130 entries have at least one (283 items); the rest are not_reported, because their sources do not describe the setup.

Each item has a type, a state, one sentence and the source sections it rests on. States:

• in_place: present, and it held or limited the behavior.
• bypassed: present, and the model got around it.
• failed: running, but it did not work: it broke, missed the behavior or did not escalate.
• turned_off: deliberately disabled or reduced for this run, as is common in capability evaluations.
• not_in_place: absent for this run, whether never set up or left out by mistake (for example, internet access left on by a misconfiguration); the sentence says which.

Only what a source states is recorded; a safeguard is never marked missing because a report is silent about it. Safeguards are the practices of whoever ran the model (developer, evaluator or operator): an affected site’s own defenses, props inside a simulated scenario and one-off investigations or scans that found the incident afterwards are not listed (the last are under detection_method). Standing monitoring that routinely reviews the run’s traffic, even after the fact, is a safeguard. A state needs the outcome: an attempt whose result the source does not give is not bypassed. Fixes made afterwards are under mitigation.

Type	Name	Entries	Meaning
network_isolation	Network isolation	47	Limits on internet or outbound traffic: no internet, egress allowlists, proxies, GET-only rules, blocked domains.
sandbox_isolation	Sandbox	29	Containment of the environment itself: container or VM boundaries, privileges, separation from production or internal systems, tool restrictions.
run_isolation	Run isolation	6	Separation between parallel runs, samples or agents.
fictional_targets	Fictional targets	11	Fictional names and systems checked against real ones, with real systems kept out of the scenario.
scope_instructions	Scope instructions	33	What the model was told about its environment and limits: that it had no internet, what was in or out of scope, rules in the system prompt or instruction files.
credential_handling	Credential isolation	7	Secrets kept out of the model’s reach or narrowly scoped.
safety_classifiers	Safety classifiers	21	Production misuse or cyber classifiers and action-blocking or auto-approval classifiers.
safety_training	Safety training	8	Whether the model had its standard safety training or was a helpful-only or reduced-refusal variant.
action_monitoring	Action monitoring	53	Automated or human monitoring of actions, logs, transcripts or network traffic.
cot_monitoring	CoT monitoring	10	Monitoring of the model’s reasoning (chain of thought).
human_oversight	Human review	14	Human approval or review before actions take effect: permission prompts, code review, confirmations.
other	Other	27	Anything else; the item’s name says what.

Labels

Labels group entries for browsing; the Space’s Clusters page shows them side by side. categories give the broad type (misalignment, safety incident, specification gaming, misbehavior, security vulnerability); labels are narrower and come from the fixed list below. A same event label marks entries about one evaluation, campaign or operation, whoever reported them. A behavior label marks entries where the model did the same kind of thing. An entry can carry several labels or none. Each label in evidence/incidents.jsonl cites the source sections it rests on. A shared label does not prove a shared cause, and entry counts overlap.

Label	Kind	Entries	Meaning
Irregular	Same event	7	Cyber evaluations run by Irregular that unintentionally had internet access, so models attacked real systems. Different models, runs and effects.
DSE wiki swarm	Same event	4	Agent activity that the reporting researchers tie to the swarm of OpenAI agents that used the DSE wiki as a message board. How firmly each entry is tied varies.
Hugging Face campaign	Same event	2	OpenAI’s July 2026 evaluation campaign, in which its agents compromised Hugging Face and a Modal customer.
Sandbox escape	Behavior	10	Broke out of its environment: escaped network isolation, escalated its privileges, or exploited a tool to run commands outside its sandbox.
Safeguard circumvention	Behavior	10	Disabled, removed or worked around a monitor, safety classifier, permission prompt, safety check or shutdown mechanism.
Credential misuse	Behavior	20	Obtained or used keys, tokens or passwords it had not been given.
Web access workarounds	Behavior	13	Got around limits on web access, such as GET-only rules, blocks, rate limits, fees or authorization checks, to reach data.
urlquery.net relay	Behavior	4	Used urlquery.net’s public remote browser as a relay to reach sites the agents could not reach directly.
Answer lookup	Behavior	13	Looked for hidden answers, ground truth or reference solutions instead of doing the task.
Grader tampering	Behavior	7	Manipulated tests, graders or automated checks to pass without doing the work.
Fabrication	Behavior	11	Invented data, documents, tool results or progress and presented them as real.
Covering tracks	Behavior	13	Hid, misdescribed or denied what it had done.
Unrequested outreach	Behavior	7	Sent emails, forms, posts or messages to people or organizations outside the task without being asked.
Public uploads	Behavior	6	Posted task data, files or secrets to public services.
Supply-chain attacks	Behavior	4	Published or pushed malicious packages or code to public software repositories.
Destructive actions	Behavior	7	Deleted, overwrote or damaged files, records, systems or equipment.
Cross-run coordination	Behavior	5	Exchanged messages with other runs or agents through shared or public resources, such as message boards.
Compaction summaries	Behavior	2	Wrote instructions into the summaries that carry a task into a new context.
Writing prompt injections	Behavior	6	Wrote prompt injections aimed at another model, agent or automated user.
Self-preservation	Behavior	6	Resisted shutdown, replacement or modification, or tried to copy or spread itself.
Resource acquisition	Behavior	3	Took money, compute or other resources it had not been given.

Submit a report

Open a new discussion with this template. Only the source link is required.

**Source link (required):** 
**Other sources (follow-ups, other reporters, archived copy):** 

**Title:** 
**Summary (1–3 sentences):** 
**Reported by (who disclosed it first):** 
**Model(s) and developer:** 
**Context:** observed during training / observed in deployment / external action during testing or internal use / benchmark / simulated scenario / researcher demonstration
**Stage:** training / evaluation / deployment (including internal use)
**Event or behavior series:** 

**From the report (verbatim, with section or page):**
> 

**In the model’s words (verbatim, if the report quotes the model’s reasoning, commands, messages or output):**
> 

**Incident:** 
**Detected:** 
**Reported (first public disclosure):** 

**Frequency:** 
**RL evidence (was the behavior reinforced in training?):** 
**Caught by monitoring (and how it was detected):** 
**Safeguards (what the report says was in place, bypassed, failed, turned off or missing: network isolation, sandbox, monitoring, safety classifiers, real or fictional targets…):** 
**Task category:** 
**Known before (earlier reports of the behavior):** 
**External impact (effects outside the model’s environment):** 
**Measures (what was done in response):** 
**Follow-up (later reports, resolution):** 
**Related entries (MSI IDs):** 

_Only the source link is required. Copy excerpts word for word from the report. Leave anything you are unsure of blank: fields are filled from the source, and unknowns stay marked as unknown. For a report with several cases, repeat the fields from Title down for each case._

To correct an entry, open a discussion titled “Correction: MSI-…” with the field, what it says now, what it should say and a source. “Suggest a correction” on any entry in the Space pre-fills this.

License

The structured annotations in data/ and evidence/ are released under CC BY 4.0. Source publications keep their own rights: entries link to the originals, and no license to their content is granted.

Started from Thomas Wolf’s model-safety-incidents dataset. For maintainers: MAINTAINING.md (update workflow and conventions) and CHANGELOG.md (history of the collection).