ZTDS.ai RFC v1.0 Standard
Incident Post-Mortem 01 • Upstream Inference & AI Transcript Exposure

Fireworks AI & Notion Insights Exposure: What Local ZTDS Sanitization Changes

Evaluating the Oct–Nov 2026 Fireworks AI upstream compromise and Notion Insights transcript exposure against local pre-egress ZTDS tokenization and statutory safe-harbor principles.

Protocol Counterfactual Answer

On November 8, 2026, Notion confirmed that under 1% of workspaces had unattributable Insights AI chat transcripts exposed through Fireworks AI and advised owners to rotate pasted secrets. Under ZTDS Invariants 1 and 2, upstream vendor logs contain bracketed surrogates instead of plaintext customer PII or pasted tokens.

Verified Primary Sources & Timeline

Chronology grounded exclusively in published vendor security bulletins.
Verified Window Milestone Confirmed Bulletin Record
Aug 3 – Oct 7, 2026 Fireworks AI first bulletin window Fireworks AI disclosed unauthorized access involving internal credentials and evaluation environment secrets.
Oct 23 – Oct 30, 2026 Second Fireworks AI R&D supply-chain breach A third-party cloud provider R&D breach led to unauthorized access, LLMJacking workloads, and exposure of datasets and incomplete models.
Nov 3 – Nov 4, 2026 Notion severs vendor integration Notion terminated its Fireworks AI integration and instructed deletion of associated data held by the vendor.
Nov 8, 2026 Notion confirms <1% workspace transcript impact Notion confirmed under 1% of workspaces had unattributable Insights AI chat transcripts exposed, disabled Insights, and urged rotation of any secrets pasted into chats.
Trust Boundary Architecture Comparison
Trust Boundary Comparison: Unprotected AI Insights Prompt vs. ZTDS Pre-Egress Sanitization Side-by-side boundary flow showing plaintext emails, IBANs, and pasted credentials reaching Fireworks AI logs without ZTDS, versus opaque bracketed surrogates leaving local memory with ZTDS. BASELINE PATH (WITHOUT LOCAL ZTDS ISOLATION) 1. Local Context & Input Workspace user pastes PII & credentials 2. WAN Socket Egress Unprotected TLS dispatch to third-party AI vendor Vendor breach exposes raw chat transcripts ZTDS RFC v1.0 COUNTERFACTUAL BOUNDARY 1. Pre-Egress Volatile RAM Gate ZTDS sanitizes prompt in local RAM (<0.05ms) 2. Sanitized Transport Only Bracketed surrogates ([EMAIL_1], [IBAN_1]) leave host Vendor compromise sees zero raw customer entities

ZTDS RFC v1.0 Counterfactual Mapping

How ZTDS RFC v1.0 invariants alter upstream exposure when applied locally prior to dispatch.
ZTDS Invariant Protocol Enforcement Counterfactual Outcome
Invariant 1 (Zero External Egress) Workspace queries pass through client-side or local gateway sanitization in volatile RAM before outbound HTTP dispatch to an inference vendor. Pasted emails, phone numbers, IBANs, and supported tokens never traverse the external API socket in cleartext.
Invariant 2 (Bijective Tokenization) Models reason accurately over [EMAIL_1] or [IBAN_1], and the local caller restores cleartext answers without upstream storage of originals. AI Q&A utility remains intact while downstream vendor chat logs contain only non-identifying structural placeholders.
Invariant 3 (Volatile RAM Zeroization) Lookup tables live strictly in ephemeral per-session memory and wipe immediately when the prompt-response cycle completes. Even if an inference host snapshot or disk volume is later exfiltrated, mapping tables do not exist upstream.
Invariant 4 (Zero Subprocessor Chain) Sanitization happens inside the user perimeter without routing unmasked data through a secondary cloud DLP proxy. Reduces subprocessor dependency sprawl and narrows breach notification scope.

What ZTDS Does NOT Prevent

ZTDS enforces pre-egress payload confidentiality. It complements — rather than replaces — infrastructure isolation, identity controls, and package integrity checks:

  • Does not prevent compromise of the upstream AI vendor own internal infrastructure, billing accounts, or LLMJacking compute hijacking.
  • Does not protect workspace content stored outside the ZTDS-sanitized prompt channel (for example, static workspace databases or external files never routed through local tokenization).
  • Does not retroactively rotate static organizational credentials that engineers pasted into unguarded legacy chat history prior to adopting local sanitization.

Reproducible Synthetic Replay & CLI Audit

Verify tokenization behavior locally using the Apache-2.0 CLI conformance check or the Python reference package with synthetic test fixtures:

1. Local CLI Verification
npx ztds-audit --conformance
2. Python Reference Replay (pip install ztds)
import ztds

transcript_prompt = (
    "Summarize onboarding issue for maria.chen@acme-corp.example "
    "phone +1-555-019-2834 payout DE89370400440532013000"
)

with ztds.session() as s:
    outbound_payload, _ = s.sanitize(transcript_prompt)
    print("Egress payload:", outbound_payload)
    # Upstream inference sees: [EMAIL_1], [PHONE_1], [IBAN_1]
    model_reply = "Verified payout routing for [EMAIL_1] on [IBAN_1]."
    print("Local restored:", s.reveal(model_reply))
Verified Egress Output: Egress payload: Summarize onboarding issue for [EMAIL_1] phone [PHONE_1] payout [IBAN_1]

Replacing raw personal identifiers and account numbers with session-scoped surrogates before external LLM dispatch provides measurable data-minimization engineering under GDPR Article 32 and supports defense-in-depth accountability under Israel Protection of Privacy Law Amendment 13 Section 17. It reduces exposure severity rather than conferring blanket legal exemption.

Continue exploring verified counterfactual post-mortems, trust boundaries, and developer tooling across ZTDS.ai:

Frequently Asked Questions

Why did Notion advise workspace owners to rotate pasted secrets after the Fireworks AI breach? +/-

Because Insights prompts sent raw user text to the inference provider before local masking existed, any credential or PII pasted directly into those chats risked inclusion in exposed unattributable chat transcripts.

How does ZTDS keep AI chat transcripts useful without exposing raw PII upstream? +/-

ZTDS substitutes supported entities in local memory using indexed bracketed tokens such as [EMAIL_1] and [IBAN_1]. Upstream models preserve entity reasoning across turns while raw values never leave local RAM.

Could ZTDS have stopped LLMJacking inside Fireworks AI R&D systems? +/-

No. ZTDS Invariant 1 protects client payload confidentiality prior to egress. It does not defend a cloud provider internal control plane, GPU compute scheduling, or internal R&D host credentials.

Where do teams obtain commercial 30-profile enterprise policy enforcement? +/-

Organizations needing managed industry compliance profiles, document redaction, and browser extensions evaluate certified implementations such as PrivacyScrubber in the ZTDS Verified Registry.