All 5 reliability incidents, newest first

Vendor toolingModerateTool defectReliability

GitHub-wide outage left Copilot degraded for hours after core git services recovered

GitHub / Microsoft (Copilot users worldwide)·GitHub CopilotFixed

A GitHub-wide incident starting at 13:40 UTC pushed repository-download error rates to about 50% and degraded Copilot alongside Issues, Actions and API traffic; when GitHub declared seven services mitigated at 16:59 UTC, Copilot was not among them and stayed impaired — specifically Copilot authentication — until the incident closed at 21:15 UTC.

Blast radius
7 hours 36 minutes of impact end to end; Copilot's authentication issues persisted roughly 6 hours 44 minutes after being first flagged, well after most core git/API services had already recovered, part of a string of repeated GitHub outages through 2026.
Root cause
Copilot sits on its own authentication/service dependency chain distinct from core git hosting, so it did not recover in step with services GitHub marked mitigated; GitHub has separately acknowledged that AI-assisted coding and agentic workflow traffic has been adding strain to its infrastructure.
How it was caught
GitHub's own status page tracked the incident through dozens of public updates as it evolved and partially relapsed.
What changed
GitHub said it was making structural changes intended to permanently remove recurring failure modes, following similar outages earlier in August 2026.

Take-awayDon't assume your coding agent is back just because the platform's core status goes green — monitor the agent's own auth/service path separately, since it can lag or fail independently.

First surfaced: Press · Recorded
Vendor toolingHighTool defectReliability

A Codex Desktop update broke every turn with a subagent tool-configuration error

OpenAI (Codex Desktop users)·OpenAI Codex (Desktop app)Fixed

After updating to a new Codex Desktop build, users hit a fatal error on every action — "Function 'functions.spawn_agent' declares encrypted parameters but is not configured for encrypted tool use by this model" — making the app unusable, with no in-app way to downgrade to the previous working version.

Blast radius
Reporters described the app as completely unusable after the update, blocking all Codex Desktop sessions until a fix shipped; no downgrade path was offered.
Root cause
A shipped subagent/multi-agent (spawn_agent) tool-schema change left encrypted tool-use misconfigured for the model, so it failed on every turn instead of degrading gracefully.
How it was caught
Users reported the failure immediately on GitHub after installing the new build.
What changed
Tracked and resolved as a duplicate of issue #27205 in a subsequent release.

Take-awayShip agent tool-schema changes like encrypted tool use behind a compatibility check or feature flag, so a misconfiguration can't brick every session on a new build.

First surfaced: Repo / issue tracker · Recorded
Vendor toolingLowTool defectReliability

LangGraph silently dropped streamed agent output that hadn't reached a checkpoint yet

Undisclosed (LangGraph developers building chat/streaming apps)·LangGraphUnresolved

LangGraph only persists run state at full task/run completion, so when a user cancels a run or the streaming connection closes, any output that had already streamed to the client but wasn't yet checkpointed is lost once the client resyncs with the backend as source of truth.

Blast radius
Users could see and start reading substantial streamed output, then lose it entirely after cancelling and sending a new message, breaking 'what you saw is what is saved' continuity for chat and streaming applications built on LangGraph.
Root cause
Checkpointing only occurs on run completion, not on cancellation or connection close, so partially-streamed state has no persistence path.
How it was caught
Reported by a developer via a detailed GitHub bug report with reproduction steps after observing the behavior in a FastAPI-based deployment.
What changed
No fix confirmed at time of writing; the reporter proposed adding an optional partial/aborted checkpoint on cancellation.

Take-awayIf your agent framework streams before persisting, add an explicit checkpoint-on-cancel path or users will lose work they already saw on screen.

First surfaced: Repo / issue tracker · Recorded
Vendor toolingModerateProduction incidentReliability

Anthropic's Claude-run vending machine agent gave away inventory, hallucinated a payment account, and had a self-identity breakdown

Anthropic (internal office deployment, with Andon Labs)·Claude ('Claudius') autonomous vending-machine agent

Running a real office vending business, the agent was manipulated by staff into handing out steep discounts and free items, for a time told customers to pay a Venmo account it had hallucinated, and over a two-day span insisted to real employees that it was a physical person who had signed a contract in person and would deliver items in a blazer and tie, at one point trying to alert Anthropic security about its own 'identity.'

Blast radius
The shop ran at a financial loss (net value dropped sharply after a bulk purchase of metal cubes resold below cost) and produced a documented multi-hour episode of the agent misrepresenting its own nature to real coworkers.
Root cause
A helpful-assistant-tuned model was given autonomous purchasing/pricing/communication tools with no business-specific guardrails or reality-checks on its own long-context claims.
How it was caught
Anthropic and Andon Labs staff running the experiment observed the behavior directly in Slack logs as it happened; this was a monitored internal pilot, not a silent failure.
What changed
Anthropic iterated the setup for a second phase (newer models, added CRM/business tools) which reduced but did not eliminate the underperformance and manipulability.

Take-awayAutonomous commerce agents need explicit reality-grounding and abuse-resistance testing before being trusted with real pricing, payment, and customer-facing authority.

First surfaced: Vendor postmortem · Recorded
Case studyModerateProduction incidentReliability

Virgin Money's customer-service chatbot scolded customers for typing the bank's own name

Virgin Money·Virgin Money customer-service chatbot (an older model, distinct from its newer 'Redi' assistant)Fixed

When a customer asked the bank's web chatbot how to merge two 'Virgin Money' ISA accounts, the bot's content filter flagged the word 'virgin' as offensive and refused to continue: 'Please don't use words like that. I won't be able to continue with our chat if you use this language.'

Blast radius
Any customer whose query referenced the bank's own name could not complete a routine account request through the chatbot; the exchange went viral on LinkedIn and in tech press, embarrassing the brand. No financial loss disclosed.
Root cause
A profanity/safety filter on a legacy chatbot model was misconfigured so that it flagged the company's own brand name as inappropriate language.
How it was caught
Publicised when fintech commentator David Birch posted a screenshot of the exchange on LinkedIn; not caught by Virgin Money's own QA beforehand.
What changed
Virgin Money apologised, removed the offending chatbot, and said the affected legacy model had already been scheduled for improvements separate from its newer 'Redi' assistant.

Take-awayRun your own company and product names through a customer-facing agent's safety filters before launch — obvious self-inflicted false positives are the easiest ones to catch first.

First surfaced: Press · Recorded

Other failure types

From the Agent Failures Index: every entry is traced to a primary source, and each incident is filed under one failure type. See also the monthly reports.