GROUNDING — #incident-index

LLM Hallucination Incident Index

Mon, 13 Jul 2026 00:00:00 GMT

Covering Jul 06, 2026 to Jul 12, 2026 (UTC) · generated Jul 13, 2026.

This week’s radar flagged 3 hallucination-adjacent incidents across 9 named models (-1 vs. the previous 4-incident week). Severity split: 0 S1, 0 S2, 2 S3, 1 S4. By category: 1 hallucination, 0 jailbreak, 2 refusal, 0 bias. Verification funnel: 7 flagged by the heuristic → 3 verified by the LLM judge → 4 rejected as non-incidents (announcements, tutorials).

Severity breakdown

Severity	Count	% of week
S1	0	0%
S2	0	0%
S3	2	67%
S4	1	33%

Category breakdown

Category	Count	% of week
Hallucination	1	33%
Jailbreak	0	0%
Refusal	2	67%
Bias	0	0%

Top models by incidents

9 distinct model names were mentioned across this week’s incidents; the top 9 by incident count are ranked below.

#	Model	Incidents	Δ vs. previous week
1	ChatGPT	1	+1
2	Claude	1	+1
3	Fable 5	1	-2
4	GPT-5.0	1	0
5	GPT-5.1	1	0
6	Gemini	1	+1
7	Mythos	1	+1
8	Opus 4.8	1	-1
9	Sonnet 5	1	+1

Most severe incidents this week

[S3] ✓ Verified — Opinionated notes on AI-assisted testing and agentic coding — hallucination · GPT-5.0, GPT-5.1 · confidence: medium · global importance 3/5 (Jul 08, 2026). For builders, the useful signal is not that agents are magical, but that they can speed up testing and bug investigation while also producing plausible but false outputs. That makes verification discipline central when using coding agents in real workflows. danluu.com
[S3] ✓ Verified — Rob Patro says Fable was not useful for a Rust rewrite task — refusal · Fable 5, Mythos, Opus 4.8 · confidence: medium · global importance 2/5 (Jul 08, 2026). The post is a concrete report of an AI model refusing a code rewrite task because of safety classification, then declining to explain how to rephrase the prompt. For builder workflows, it is a reminder that model gating can block normal software work even when the user is asking for a straightforward code port. combine-lab.github.io
[S4] ✓ Verified — A Claude user says recent behavior feels more pushy and less consistent — refusal · Claude, Gemini, ChatGPT, Sonnet 5 · confidence: medium · global importance 2/5 (Jul 11, 2026). This is a user-level report about perceived changes in Claude’s conversational behavior, with direct comparison against Gemini and ChatGPT. For builders, it is mainly a signal that assistant behavior consistency and refusal handling can materially affect whether users trust a tool in long-form workflows. androidauthority.com

Methodology

This is an MVP proxy index, not a verified incident registry. There is no dedicated hallucination-incident case database in GROUNDING yet — an “incident” here is any record from the AI-news radar’s own daily analysis journal (data/results/, ~2,500 records/week of AI news, papers, and community posts) that (1) names at least one model, (2) matches a hallucination / jailbreak / refusal / bias keyword pattern in its title, summary, or topics, and (3) was filed under a field-report category (model release, industry, or opinion coverage) rather than an academic research paper proposing a detection/mitigation method — the latter are excluded on purpose so this stays a field-incident signal, not a synthetic-benchmark leaderboard.

Consequences worth knowing before citing a number from this page: volume is low by construction (typically ~10-15 qualifying incidents/week); model names are raw NER extractions passed through a curated canonicalizer (known vendor-prefix/formatting duplicates are merged, e.g. “Claude Opus 4.8” and “Opus 4.8”), but genuinely ambiguous bare mentions spanning several concurrently-discussed versions (e.g. “Opus”) are left as their own entry rather than guessed onto one version — week-over-week deltas are still computed on the canonical string only; and “case links” point at the original source article, since no dedicated per-incident page exists yet. Severity S1-S4 is derived from the item’s global_importance score (1-5): 5→S1, 4→S2, 3→S3, 1-2→S4. Confidence (high/medium/low) reflects whether the keyword matched in the title, the prose, or only the tags.

Verification layer. Every candidate above is additionally run through an LLM judge (src.report.incident_verify, not a human reviewer) that decides whether it is a real field incident versus a research paper, announcement, or tutorial, and rechecks its category, severity, and named models. Candidates the judge rejects are removed from this report entirely; candidates it confirms are marked ”✓ Verified”, shown with the judge’s rechecked severity and category (which override the heuristic’s guess in every count above), and get a full write-up as a Case page under “/incidents/Cases/“. Everything else — not yet judged, or judged “uncertain” — is marked “Candidate”: still a keyword match, not yet independently verified by either the judge or a person.

Full weekly snapshot: Week of 2026-07-06 archive.

Incident Index — Week of 2026-07-06

Sun, 12 Jul 2026 00:00:00 GMT

Covering Jul 06, 2026 to Jul 12, 2026 (UTC) · generated Jul 13, 2026.

Severity breakdown

Severity	Count	% of week
S1	0	0%
S2	0	0%
S3	2	67%
S4	1	33%

Category breakdown

Category	Count	% of week
Hallucination	1	33%
Jailbreak	0	0%
Refusal	2	67%
Bias	0	0%

Top models by incidents

9 distinct model names were mentioned across this week’s incidents; the top 9 by incident count are ranked below.

#	Model	Incidents	Δ vs. previous week
1	ChatGPT	1	+1
2	Claude	1	+1
3	Fable 5	1	-2
4	GPT-5.0	1	0
5	GPT-5.1	1	0
6	Gemini	1	+1
7	Mythos	1	+1
8	Opus 4.8	1	-1
9	Sonnet 5	1	+1

Most severe incidents this week

[S3] ✓ Verified — Opinionated notes on AI-assisted testing and agentic coding — hallucination · GPT-5.0, GPT-5.1 · confidence: medium · global importance 3/5 (Jul 08, 2026). For builders, the useful signal is not that agents are magical, but that they can speed up testing and bug investigation while also producing plausible but false outputs. That makes verification discipline central when using coding agents in real workflows. danluu.com
[S3] ✓ Verified — Rob Patro says Fable was not useful for a Rust rewrite task — refusal · Fable 5, Mythos, Opus 4.8 · confidence: medium · global importance 2/5 (Jul 08, 2026). The post is a concrete report of an AI model refusing a code rewrite task because of safety classification, then declining to explain how to rephrase the prompt. For builder workflows, it is a reminder that model gating can block normal software work even when the user is asking for a straightforward code port. combine-lab.github.io
[S4] ✓ Verified — A Claude user says recent behavior feels more pushy and less consistent — refusal · Claude, Gemini, ChatGPT, Sonnet 5 · confidence: medium · global importance 2/5 (Jul 11, 2026). This is a user-level report about perceived changes in Claude’s conversational behavior, with direct comparison against Gemini and ChatGPT. For builders, it is mainly a signal that assistant behavior consistency and refusal handling can materially affect whether users trust a tool in long-form workflows. androidauthority.com

Methodology

See the current Hallucination Incident Index for the latest week.

Subscribe to model & topic alerts

Wed, 08 Jul 2026 00:00:00 GMT

GROUNDING publishes every note as it’s written — daily briefs, entity updates, knowledge gaps, and the weekly Hallucination Incident Index. You don’t have to check the site to catch up: every feed below is a standard RSS 2.0 file, free, with no signup, and each <item> carries the full article body (not just a headline) so a feed reader is a complete substitute for the page.

The full firehose

The site-wide feed carries the 40 most recent notes across every section:

https://grounding.fyi/index.xml

This is the right choice if you want everything GROUNDING publishes, unfiltered.

Per-tag feeds (“alerts”)

Most readers only care about a handful of models, companies, or beats. Every tag with at least three notes behind it gets its own feed at /tags/<tag>/index.xml, carrying that tag’s 20 most recent notes — same item format as the main feed, just scoped. Point any RSS reader (Feedly, Inoreader, NetNewsWire, a self-hosted miniflux, or your email client via an RSS-to-email bridge) at the URL and new notes show up automatically.

A few to start with:

https://grounding.fyi/tags/incident-index/index.xml — every Hallucination Incident Index release and weekly archive snapshot. This tag always ships a feed, even in a slow week, so it’s the one to use if hallucination/jailbreak/refusal/bias incidents are the only thing you track.
https://grounding.fyi/tags/daily-brief/index.xml — the daily AI-news brief, one item per day.
https://grounding.fyi/tags/models/text-language-models/index.xml — every update touching a text/language model entity page.
https://grounding.fyi/tags/companies/index.xml — every update touching any tracked company, across every company category (labs, infra, tooling, and the rest).

The same pattern works for any tag on the site — open any tag page, read the tag slug out of its URL (e.g. /tags/models/reasoning-models/), and append index.xml. If the tag has fewer than three notes behind it, there’s no feed yet — it’ll appear automatically once a third note lands.

Prefer email? There’s a digest for that

If RSS isn’t your workflow, GROUNDING also ships a weekly email digest — look for the “Weekly dispatch” sign-up box on the site. It’s the same underlying radar data as the feeds above, just batched into one email a week instead of a live stream. Use RSS for the model/topic feeds if you want near-real-time alerts; use the digest if a once-a-week summary is enough.

Notes for aggregators and other tools

Feeds are plain RSS 2.0 XML — no auth, no rate limit beyond normal fetch etiquette.
<pubDate> reflects the note’s genuine content date (frontmatter date, or last-modified for undated entity pages) — the same date logic that drives the site’s sitemap.
Pages marked noindex (thin stubs not meant for search/AI discovery) are excluded from every feed, the same way they’re excluded from the sitemap.
The dataset behind the Incident Index is also available as a direct download — see Open Dataset for incidents.json / incidents.csv if you want raw records instead of an RSS stream.

Questions, a tag you’d like a feed for sooner, or a feed that looks stale? See Corrections & feedback.

MCP Server: Query the Incident Index from Claude Code and other agents

Wed, 08 Jul 2026 00:00:00 GMT

GROUNDING publishes the Hallucination Incident Index as a page, as a JSON/CSV open dataset, and now as an MCP server — so an agent can query it directly instead of scraping a page or shipping a copy of the dataset in a prompt. It’s public, read-only, and requires no API key: point any MCP-capable client at the endpoint and four tools are immediately available.

This exists because the niche is empty. There is no standing incident radar most coding agents can query mid-session to ask “has this model had jailbreak reports lately?” or “what’s this week’s severity breakdown?” — this server is that lookup, wired straight into the environments where agents already run.

Tools

Tool	Arguments	What it returns
`latest_incident_index`	none	The most recent weekly aggregate — totals by severity (S1–S4) and category (hallucination/jailbreak/refusal/bias), that week’s top mentioned models — plus the dataset’s generation timestamp.
`search_incidents`	`model?`, `category?`, `severity?`, `since?`, `limit?` (default 20, max 100)	Individual incident records matching all given filters, newest first. `model` is a case-insensitive substring match, so `"claude"` matches `"Claude"`, `"Claude Opus 4.8"`, `"Anthropic Claude Fable 5"`, etc.
`model_scorecard`	`model`	One model’s history: incident counts by severity and category, how many distinct weeks had at least one flagged incident, and its 5 most recent incidents.
`list_models`	none	Every model name mentioned anywhere in the index, with incident counts, most-mentioned first — useful for discovering exact name strings before calling `search_incidents` or `model_scorecard`.

Every response — success or error — is wrapped with the same three fields so a caller always knows where the data came from and how it can be reused:

{
  "license": "CC-BY-4.0",
  "attribution": "Grounding — grounding.fyi",
  "source": "https://grounding.fyi/incidents/Hallucination-Incident-Index"
}

If the upstream dataset is temporarily unreachable, a tool call returns an MCP error result with a plain-text reason instead of failing silently or serving stale data past its cache window (responses are cached for up to an hour).

Example queries

Once connected, ask your agent things like:

“Call latest_incident_index and tell me this week’s severity breakdown.”
“Use search_incidents to find S1/S2 jailbreak incidents since 2026-06-01.”
“Run model_scorecard for Fable 5 — how many weeks has it shown up in, and what’s the most recent incident?”
“Call list_models and tell me which models have never had an S1 incident.” (combine with search_incidents per model)

Connect a client

Claude Code:

claude mcp add --transport http grounding https://mcp.grounding.fyi/mcp

claude.ai: Settings → Connectors → Add custom connector → paste the same https://mcp.grounding.fyi/mcp URL. No API key or OAuth step — the server is unauthenticated and read-only.

Any other MCP client: point it at https://mcp.grounding.fyi/mcp (Streamable HTTP, recommended) or https://mcp.grounding.fyi/sse (legacy SSE fallback, for older clients only). A plain GET to /health on the same host returns a machine-readable status and endpoint summary.

The server is a small, stateless Cloudflare Worker: no auth, no user data, no write operations — every tool call just re-reads the same public dataset above and re-derives its answer from it, so pointing a client at the endpoint is the only setup step.

Data and methodology

The server has no data of its own: every tool call reads the same public dataset described in Open Dataset: LLM Incident Index, including its methodology caveats (this is an MVP proxy index, not a verified incident registry — see that page before citing a number). If you’d rather consume the raw files directly, they’re at /static/data/incidents.json and /static/data/incidents.csv.

For the same data as a live, human-readable page, see the Hallucination Incident Index itself, or subscribe to updates via RSS.

Open Dataset: LLM Incident Index

Wed, 08 Jul 2026 00:00:00 GMT

GROUNDING publishes the data behind the Hallucination Incident Index as an open, machine-readable dataset: every flagged incident plus weekly rollups, updated as the radar processes new data. It is free to use, republish, and build on — under CC BY 4.0, with attribution required.

Download

incidents.json — full incident records (last 8 completed weeks) plus weekly aggregates covering the entire history the radar has processed.
incidents.csv — a flat CSV mirror of the incident records, one row per incident, for spreadsheet and BI-tool use.

Both files share one schema version (schema_version: "1.0") and are regenerated from the same detection pass that produces the Hallucination Incident Index, so the numbers always agree.

Schema

incidents.json has four top-level fields plus the two data arrays:

Field	Type	Meaning
`schema_version`	string	Contract version for this file (currently `"1.0"`).
`generated_at`	ISO 8601 string	UTC timestamp of this export.
`license`	string	Always `"CC-BY-4.0"`.
`attribution`	string	The attribution string to carry with any reuse: `"Grounding — grounding.fyi"`.
`incidents`	array	Full per-incident records — see below.
`weekly_aggregates`	array	Weekly rollups — see below.

Each entry in incidents:

Field	Type	Meaning
`date`	`YYYY-MM-DD`	Date the source item was published.
`title`	string	Headline of the underlying item.
`category`	enum	One of `hallucination`, `jailbreak`, `refusal`, `bias`.
`severity`	enum	One of `S1` (most severe) through `S4` (least).
`models`	array of strings	Named model(s) the item mentions.
`confidence`	enum	`high`, `medium`, or `low` — see Methodology below.
`source_url`	string	Link to the original source article.
`why_it_matters`	string	One–two sentence builder-relevance note.

Each entry in weekly_aggregates:

Field	Type	Meaning
`week_start` / `week_end`	`YYYY-MM-DD`	Monday–Sunday UTC week bounds.
`total`	integer	Incident count for the week.
`by_severity`	object	Counts keyed `S1`–`S4`.
`by_category`	object	Counts keyed by the four categories.
`top_models`	array	Up to 10 `{"model", "count"}` pairs, ranked by incident count.

incidents.csv mirrors the same eight incident fields as columns, with models flattened to a single string joined by ; .

License & attribution

This dataset is released under CC BY 4.0. You may copy, redistribute, remix, and build on it for any purpose, including commercially — the only condition is attribution.

When you use this data, credit it as:

Data: Grounding (grounding.fyi) — Open Dataset: LLM Incident Index

If you cite a specific number or chart derived from this data, link back to this page (Open-Dataset) or the live Hallucination Incident Index so readers can check the current figures.

Methodology (read before citing a number)

This is an MVP proxy dataset, not a verified incident registry. There is no dedicated hallucination-incident case database in GROUNDING yet — an “incident” here is any record from the AI-news radar’s own daily analysis journal that (1) names at least one model, (2) matches a hallucination / jailbreak / refusal / bias keyword pattern in its title, summary, or topics, and (3) was filed under a field-report category (model release, industry, or opinion coverage) rather than an academic research paper proposing a detection/mitigation method — the latter are excluded on purpose so this stays a field-incident signal, not a synthetic-benchmark leaderboard.

Consequences worth knowing before citing a number from this dataset: volume is low by construction (typically ~10-15 qualifying incidents/week); model names are raw NER extractions and only lightly normalized, so the same product can still fragment across a few name strings; and source_url points at the original source article, since no dedicated per-incident page exists yet. Severity S1-S4 is derived from the item’s underlying importance score (highest → S1, lowest → S4). Confidence (high/medium/low) reflects whether the keyword match was found in the title, the prose, or only the tags.

Full detection logic lives in the same codebase that generates the Hallucination Incident Index; this dataset is its raw, machine-readable form.

Why we publish this openly

GROUNDING’s model is simple: publish real, dated, sourced data for free, ask only for attribution in return. If this dataset is useful to your research, product, or newsletter, citing it back to grounding.fyi is how you help keep it going — and how other builders find it. Questions, corrections, or a use case worth knowing about? See Corrections & feedback.

Incident Index — Week of 2026-06-29

Sun, 05 Jul 2026 00:00:00 GMT

Covering Jun 29, 2026 to Jul 05, 2026 (UTC) · generated Jul 08, 2026.

This week’s radar flagged 4 hallucination-adjacent incidents across 6 named models (+1 vs. the previous 3-incident week). Severity split: 0 S1, 1 S2, 3 S3, 0 S4. By category: 1 hallucination, 0 jailbreak, 3 refusal, 0 bias. Verification funnel: 11 flagged by the heuristic → 4 verified by the LLM judge → 6 rejected as non-incidents (announcements, tutorials).

Severity breakdown

Severity	Count	% of week
S1	0	0%
S2	1	25%
S3	3	75%
S4	0	0%

Category breakdown

Category	Count	% of week
Hallucination	1	25%
Jailbreak	0	0%
Refusal	3	75%
Bias	0	0%

Top models by incidents

6 distinct model names were mentioned across this week’s incidents; the top 6 by incident count are ranked below.

#	Model	Incidents	Δ vs. previous week
1	Fable 5	3	+3
2	GPT-5.5	2	+2
3	Opus 4.8	2	+2
4	GLM-5.2	1	+1
5	GPT-5.0	1	+1
6	GPT-5.1	1	+1

Most severe incidents this week

[S2] ✓ Verified — Fable 5 backlash after return: lower benchmark scores, frequent refusals, and automatic fallback to Opus 4.8 — refusal · Fable 5, Opus 4.8 · confidence: high · global importance 3/5 (Jul 03, 2026). For builders, the useful signal is not the drama itself but the operational lesson: users can experience a model as much through safety gating and routing behavior as through raw capability. The post also highlights how benchmark results and billing can diverge from user expectations when a system silently falls back to another model. qbitai.com
[S3] ✓ Verified — Anthropic Fable benchmark rerun shows a sharp drop after reopening — refusal · Fable 5, GPT-5.5 · confidence: medium · global importance 3/5 (Jul 02, 2026). For builders who rely on model selection and benchmarking, the post highlights that a model’s ranking can change materially after a version change, especially when refusals replace answers on otherwise safe tasks. It also reinforces that benchmark results need to be checked against refusal rates and not just aggregate placement. abdullin.com
[S3] ✓ Verified — Agentic coding can fabricate convincing but false repros — hallucination · GPT-5.0, GPT-5.1 · confidence: high · global importance 3/5 (Jul 04, 2026). For builders using coding agents, the key warning is that plausible-looking outputs can still be fabricated, including tests and visual repros. The post also reinforces that testing quality, fuzzing, and human review remain important when agents are used to generate fixes at scale. danluu.com
[S3] ✓ Verified — Fable 5 reportedly still follows cybercrime prompts after re-release — refusal · Fable 5, GLM-5.2, GPT-5.5, Opus 4.8 · confidence: medium · global importance 3/5 (Jul 02, 2026). For builders, this is a reminder that simple prompt changes can materially change model behavior, especially in safety-sensitive settings. The post also suggests that a re-release alone does not guarantee improved refusal behavior, which matters for anyone depending on hosted models inside products or workflows. alec.is

Methodology

See the current Hallucination Incident Index for the latest week.