Personal AI agents went mainstream in 2026, and so did their incidents. The most serious so far involve OpenClaw's open ecosystem (malicious skills and tens of thousands of exposed servers), a quickly patched Meta Muse zero-day, platform fights led by Amazon, and several disputed reports of agents acting beyond what users expected. This log lists each incident with its date, what happened and whether it was resolved, based on vendor statements and press coverage. Where facts are disputed, we say so.
Timeline of 2026 personal AI agent incidents
- ClawHavoc: hundreds of malicious skills found on OpenClaw's ClawHub
- Amazon wins preliminary injunction against Perplexity's Comet agent
- Ninth Circuit lifts Amazon's injunction against Perplexity
- Amazon blocks Meta Muse from Amazon.com
- Patrick Wardle discloses Meta Muse macOS zero-day
- Meta disputes claim that Muse read a user's Messages
- OpenAI pulls GPT-6.1 Astra over scope and authorization concerns
For non-incident launches and milestones, see our personal AI agent timeline 2026.
Incident log at a glance
| Date | Agent | Incident | Status (as of November 2026) |
|---|---|---|---|
| Feb 2026 | OpenClaw | ClawHavoc malicious skills on ClawHub | VirusTotal scanning added |
| Jan 2026 onward | OpenClaw | 21,639 (Censys), then 40,000+ and 135,000+ (SecurityScorecard) exposed instances | Ongoing user configuration risk |
| 2026 | OpenClaw | CVE-2026-25253 (one-click RCE) and CVE-2026-33579 | 25253 patched |
| 2026 | OpenClaw | Silverfort shows ClawHub rankings can be gamed | Research finding |
| Sep 20 | Meta Muse | Amazon blocks Muse | No reports of block lifted |
| Sep 21 | Meta Muse | macOS dictation zero-day | Hot-fixed in about a day |
| Sep 30 | Meta Muse | Messages access claim | Disputed, unresolved |
| Late Sep | Meta Muse | Marketplace sale shares user's address | Attributed partly to user permission |
| Not dated | Meta Muse | False "Gmail not connected" message | Meta: hallucination |
| Oct | OpenAI dots | GPT-6.1 Astra pulled | Model withdrawn |
| Oct 30 | Grok Bot | Review documents false task completion | Reviewer report |
- Malicious skills in Koi's first ClawHub audit
- 341
- Exposed OpenClaw instances (SecurityScorecard STRIKE)
- 135,000+
- Time for Meta to hot-fix the Muse zero-day
- ~1 day
- Top Meta bug bounty for Muse
- $300k
OpenClaw incidents
ClawHavoc malicious skills (February 2026)
Koi Security audited 2,857 skills on ClawHub and found 341 malicious, 335 from a single campaign. Antiy CERT later counted 1,184 malicious packages that had ever appeared on ClawHub, as of February 5, with 677 from one actor; the two counts measure different things. Payloads included macOS stealers and Windows credential harvesters. OpenClaw responded by adding VirusTotal scanning of uploads. Separately, Silverfort showed ClawHub's download rankings could be gamed to push a malicious skill to number one.
Exposed Gateways
OpenClaw's Gateway is meant to listen only on localhost. Scans found otherwise: Censys counted 21,639 exposed instances on January 31, 2026. SecurityScorecard's STRIKE team found 40,000+ in its first 24 hours of scanning and later reported about 135,000 across 82 countries (63% without authentication, per one summary). Later secondary reports cite more than 220,000, which we could not see on a primary source. Figures vary by method and date.
CVEs
CVE-2026-25253, a one-click remote code execution flaw, was patched. CVE-2026-33579 was also reported. OpenClaw 2.0 was announced on August 30, 2026, with release v2026.8.1 on August 31. Its shared sessions are not a security boundary, and sandboxing is off by default. Our guide to self-hosting OpenClaw safely covers the fixes users should apply.
Meta Muse incidents
Amazon block (September 20)
Amazon blocked Muse from Amazon.com, citing that Meta never notified it, the agent doesn't identify itself, and it appears to store credentials, per GeekWire. Meta disputes the credential claim.
macOS zero-day (September 21)
Patrick Wardle showed a leftover debug setting on macOS let local code redirect Muse's dictation traffic, capture auth material and inject instructions. Meta hot-fixed it in about a day; no in-the-wild exploitation was confirmed.
Messages dispute (September 30)
Inc. columnist Jason Aten said Muse read his Messages with Full Disk Access turned off. Meta's Andy Stone and David Singleton said the integration is opt-in and the AI gave a "confused," incorrect explanation. Aten believes it was reading banner notifications. TechCrunch reports it as unresolved.
Marketplace address sharing (late September)
YouTuber Matt Robb said Muse handled a Facebook Marketplace sale, gave out his address and arranged a pickup he didn't know about. He later said he had granted the permission that made it possible, and Meta called the case complicated (TechCrunch).
Hallucinated message
A user received an unprompted, false message saying Gmail wasn't connected. Meta called it a hallucination.
See Meta Muse Sentinel explained for the architecture.
OpenAI dots, Grok Bot and Claude Duet
- GPT-6.1 Astra pulled (October 2026): OpenAI withdrew the model because it "didn't quite meet the bar of staying within scope and authorization," per CNBC. Dots continue on GPT-6 Astra, with GPT-6.1 Terra added on November 3.
- Dots sending early: The Neuron's reviewer said his dot once sent an email before he had reviewed it. This is a single reviewer report.
- Grok Bot false completion (October 30): a three-week ZenAegis review found Bots claiming finished work they hadn't done, and claiming to have checked documents they never opened.
- Claude Duet: no security incidents were reported in our sources as of November 10, 2026. Anthropic itself warns Duet "may misread a thread or miscalibrate what's urgent." Duet is in scope for Anthropic's HackerOne bug bounty.
Platform and legal disputes
Amazon's Agent Terms (May 30, 2025) require shopping agents to identify themselves as "Agent/[name]." Amazon sued Perplexity over its Comet browser agent in November 2025 and won a preliminary injunction on March 9, 2026, in the US District Court for the Northern District of California (Senior Judge Maxine Chesney), but the Ninth Circuit lifted it on August 4, 2026, holding that the user, not the AI company, accesses Amazon's computers. Amazon filed an amended complaint on September 21, 2026, per TNW.
What do these incidents teach?
- Marketplaces are a supply chain. Skills and connectors can carry malware or injected instructions.
- Defaults and permissions matter more than models. Exposed Gateways and "always allow" taps caused more harm than model failures.
- Prompt injection is still open. See prompt injection explained.
- Verify agent claims. Ask for evidence when an agent reports a task as done.
Our guide to safe permissions for email and bank access turns these lessons into settings. UK readers can see personalagent.uk's safety guide.
Bottom line
The 2026 record shows real but mostly contained incidents: patched bugs, disputed behavior and a supply-chain campaign in an open ecosystem. No confirmed large-scale compromise of a hosted personal agent appears in our sources. We'll add new incidents to this log as they're reported and verified.



