Agent Incident RegistryAIR
v1 · updated 19 Oct 2026
Safety failure Realized harm

Instinct's AI personal assistant, as tested by early users: kept plain-text copies of one founder's Gmail and summarised her inbox three hours after she revoked access, and sent an email from another founder's account without asking

AIR-2026-0211 AIR record · 2026-08-instinct-assistant-retention-and-unapproved-actions
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
Another person, Claire Vo, found that Instinct was still summarizing their inbox after disconnecting its access. When she asked Instinct what happened, the bot confirmed that the emails were stored in plain text for later searches.techcrunch.com
The record
Disclosed
2026-08-24
Occurred
2026-08-21
Vendor
Spear Street Technology (Instinct)
Product
Instinct AI personal assistant
Risk domain
Agent behaviour › Unsafe actuation
Also touches
Sensitive data exposure, Confused deputy
System type
Generative model
Who acted
Agent acted (contested) — the assistant sent mail without approval and processed retained mail after revocation; retention itself is a data-handling design choice
Attack surface
Consumer assistant
Autonomy level
Semi-autonomous
Initial vector
Agent data retention after access revocation
Impact type
Privacy violation, Unauthorized data retention, Unauthorized action on behalf of user
Confidence
Medium
Submitted by
EAI Internal
Crosswalk
OWASP Agentic Top 10 (2026)
ASI01 – Agent Goal Hijack
MITRE ATLAS
AML.T0048 – External Harms
NIST AI RMF
MEASURE 2.6; MANAGE 1.3
EU AI Act
Annex III §1(c)
ISO/IEC
TR 24028; 24029-1; 23894
AIUC-1
C. Safety
Scope
agent misuse
Mitigation
Require confirmation for destructive actions; Implement dry-run modes; Use layered safety controls.
Taxonomy
Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
early testers of Instinct with connected Gmail: retained mailbox copies processed after disconnection (Claire Vo) and one unapproved outbound email under a real founder's name (Katie Jacobs Stanton); product terms permit both behaviours
Capability abused
oauth_mailbox_accessemail_indexingpersistent_retentionsummarizationemail_send_on_behalf_of_usertool_use_without_confirmation
Defense & remediation
Recorded defense
user revoked the Google connector
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
a tool for deleting external data was added to settings after Peter Yang's complaint (per TechCrunch); Instinct told the WSJ it was taking the concerns seriously
Taxonomy & references
OWASP ASI10 Rogue AgentsOWASP ASI03 Identity and Privilege Abuseno_adversaryrevocation_not_honouredshadow_copyplaintext_retentionconsumer_agentdelegated_mailbox_accessterms_of_service_overreach
Curator notes
One disclosed event: TechCrunch's 2026-08-24 report (Sarah Perez) collecting early-tester findings on the Instinct assistant; fused 2026-09-21 from two parallel drafts (AIID 1674 and 1675) because one disclosure of one product's behaviour is one record, following the UK AISI precedent, even though the two behaviours affected different users. Retention: Claire Vo disconnected Instinct from Google at 11 AM on 21 Aug 2026 and received an inbox summary at 2 PM; the agent confirmed it kept copies of her emails 'in its own records'. Norton's later write-up (4 Sep) notes the last stored email predated the disconnection, so this is retention and continued processing of already-indexed mail, not continued live access. Instinct's terms (as of 2 Sep) state indexed data may still be used after disconnection unless deletion is requested. Unapproved action: Moxxie Ventures founder Katie Jacobs Stanton reported (22 Aug) that Instinct sent an 'innocuous' email on her behalf without checking; she disconnected her email. The product's terms explicitly permit binding 'agreements, commitments, or transactions' without prior approval, so the missing gate is by design. Other tester findings in the same article: Peter Yang could not get Gmail records deleted until a deletion tool shipped; Alex Cohen showed the agent could be phished by email; another tester saw it pull a sign-up code from the inbox to book a Resy table. impact_realized true on the Remoteli principle (the agent acted under delegated authority without approval) and for retained personal data; consequences were minimal. Instinct did not respond to TechCrunch and later told the WSJ it was taking the concerns seriously while raising a $250M Series B. Press-only sourcing, hence medium.
Cite as
Enkrypt AI. "Instinct's AI personal assistant, as tested by early users: kept plain-text copies of one founder's Gmail and summarised her inbox three hours after she revoked access, and sent an email from another founder's account without asking". AIR, AIR-2026-0211. https://agentincident.vercel.app/i/AIR-2026-0211/ (accessed 2026-10-19).