Skip to navigation

Threat reportThreat IntelligenceTL-2026-2915

Coordinated model-distillation campaign against OpenAI: 15,000+ accounts attempt to extract protected model reasoning, linked to Moonshot AI-associated individuals

mediumMONITORING

OpenAI disrupted a coordinated adversarial model-distillation effort aimed at extracting protected internal reasoning from its models. Activity began at low volume on July 1, 2026, spiked July 24-25 (~16,000 extraction-pattern requests from 4,000+ users), and a wider cluster of 15,000+ users was fully disrupted by July 28, 2026.

Key facts

IdentifiersTL-2026-2915
Also known asOpenAI reasoning-extraction campaign, Encrypted reasoning replay distillation
SeverityMEDIUM
Affected productsOpenAI OpenAI model APIs (protected/encrypted reasoning)
Nation-state nexusChina
MITRE ATT&CK / ATLAS techniquesAML.T0005 Create Proxy AI Model, AML.T0024 Exfiltration via AI Inference API, AML.T0024.002 Extract AI Model, AML.T0040 AI Model Inference API Access, AML.T0051 LLM Prompt Injection, T1585.003 Establish Accounts: Cloud Accounts
MotivationFINANCIAL
Target sectorstechnology, artificial-intelligence
Target regionsGlobal
Detection rules9 (1 shown free)
Sources8 cited sources

Detection coverage

As of 2026-10-04, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-2915 across Splunk SPL, Microsoft KQL and Sigma, covering 4 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: Shared Prompt Across Many New LLM Accounts (Account-Farm Cluster) — Post-Exploitation (SPL)

Identifies one prompt/context fingerprint reused by 25+ distinct accounts, mostly newly created, within 24 hours, and then adds a second view of newly created accounts that burst-signup from a shared IP/ASN or email domain and immediately start inference calls. The campaign linked a 15,000+ account cluster through reused attacker-controlled prompts and bulk fraudulent signups via third-party services (Establish Accounts: Cloud Accounts, then Valid Accounts: Cloud Accounts to use them); this surfaces the same account-farm signature in your own LLM gateway or AI-enabled product.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1585, T1585.003, T1078.004, T0040
  • Data sources: llm_gateway request logs (index=llm_gateway sourcetype=llm:gateway:request), llm_gateway signup logs (index=llm_gateway sourcetype=llm:gateway:signup), LLMGatewayRequests_CL (Sentinel custom table), identity provider / signup logs (account_age_days, signup_email_domain)
  • Severity: medium
  • Confidence: medium

Splunk SPL

index=llm_gateway sourcetype="llm:gateway:request" earliest=-24h@h
| eval prompt_sha256=coalesce(prompt_sha256, md5(substr(request_body, 1, 2000)))
| eval new_account=if(tonumber(account_age_days)<=7, 1, 0)
| stats count as requests dc(user_id) as accounts sum(new_account) as new_account_reqs dc(src_ip) as src_ips dc(src_asn) as src_asns dc(signup_email_domain) as email_domains values(signup_email_domain) as top_email_domains earliest(_time) as first_seen latest(_time) as last_seen by prompt_sha256
| eval new_account_ratio=round(new_account_reqs/requests, 2)
| where accounts>=25 AND requests>=100 AND new_account_ratio>=0.5
| append [ search index=llm_gateway sourcetype="llm:gateway:signup" earliest=-24h@h
    | stats dc(user_id) as signups values(signup_email_domain) as signup_domains by src_ip src_asn
    | where signups>=20
    | eval prompt_sha256="signup-burst:".src_ip ]
| convert ctime(first_seen) ctime(last_seen)
| sort - accounts

Microsoft KQL

let PromptFarm = LLMGatewayRequests_CL
| where TimeGenerated > ago(24h)
| extend PromptSha256 = coalesce(PromptSha256_s, hash_sha256(substring(RequestBody_s, 0, 2000)))
| extend NewAccount = iff(toint(AccountAgeDays_d) <= 7, 1, 0)
| summarize Requests = count(), Accounts = dcount(UserId_s), NewAccountReqs = sum(NewAccount), SrcIps = dcount(SrcIp_s), SrcAsns = dcount(SrcAsn_s), EmailDomains = make_set(SignupEmailDomain_s, 10), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated) by PromptSha256
| extend NewAccountRatio = round(todouble(NewAccountReqs) / Requests, 2)
| where Accounts >= 25 and Requests >= 100 and NewAccountRatio >= 0.5
| extend Signal = "shared-prompt-cluster"
| project FirstSeen, LastSeen, Signal, PromptSha256, Accounts, Requests, SrcIps, SrcAsns, EmailDomains, NewAccountRatio;
let SignupBurst = LLMGatewaySignups_CL
| where TimeGenerated > ago(24h)
| summarize Signups = dcount(UserId_s), EmailDomains = make_set(SignupEmailDomain_s, 10), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated) by SrcIp_s, SrcAsn_s
| where Signups >= 20
| extend Signal = "signup-burst", PromptSha256 = strcat("signup-burst:", SrcIp_s), Accounts = Signups
| project FirstSeen, LastSeen, Signal, PromptSha256, Accounts, EmailDomains;
union kind=outer PromptFarm, SignupBurst
| order by Accounts desc

Sigma

title: LLM API Request From Recently Created Account
name: llm_api_request_new_account
id: 7b2e9a14-5c36-4d08-a1f7-2e8c4d90b365
status: experimental
description: Base rule matching LLM gateway requests from accounts created within the last 7 days. Used by the correlation rule below.
author: Threadlinqs Team
date: 2026-10-04
logsource:
  category: application
  product: llm_gateway
  service: request
detection:
  selection:
    account_age_days|lte: 7
  condition: selection
level: informational
---
title: Shared Prompt Fingerprint Across Many New LLM Accounts (Distillation Account Farm)
id: c19d4e87-02fa-4a63-b5d1-9e3a7f6c2081
status: experimental
description: Detects 25 or more distinct recently created accounts sending the same prompt fingerprint within 24 hours, the account-farm signature of the 2026 model-distillation campaign (15,000+ accounts sharing an extraction prompt pattern, bulk fraudulent signups).
references:
  - https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/
  - https://anthropic.com/news/detecting-and-preventing-distillation-attacks
author: Threadlinqs Team
date: 2026-10-04
tags:
  - attack.resource_development
  - attack.t1585
  - attack.t1585.003
  - attack.initial_access
  - attack.t1078.004
  - atlas.aml.t0040
correlation:
  type: value_count
  rules:
    - llm_api_request_new_account
  group-by:
    - prompt_sha256
  timespan: 24h
  condition:
    gte: 25
    field: user_id
falsepositives:
  - Product launches or classroom/hackathon onboarding where many new users run the same template prompt
  - Shared SDK quickstart snippets executed by new developers
level: medium

Known false positives:

  • Product launches, classroom or hackathon onboarding where many new users run the same template prompt
  • Shared SDK quickstart snippets executed by new developers
  • Corporate NAT or university egress IPs producing legitimate signup bursts

Tuning notes: Enrich gateway events with account_age_days, signup_email_domain and src_asn from your IdP/billing/GeoIP system. Exclude your own onboarding/quickstart prompt hashes via a lookup. Lower accounts>=25 for small tenants; raise it for public-facing products. Pivot on src_ip/ASN and signup_email_domain from hits to scope the full cluster, since the reported cluster was linked by shared prompts and third-party signup infrastructure. The signup-burst branch is optional - drop it if signup events are not ingested.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Encrypted Reasoning Replay With Decrypt/Transcribe Prompt — Exploitation (KQL) · unlock with TL Creds
  • Shared Prompt Across Many New LLM Accounts (Account-Farm Cluster) — Post-Exploitation (KQL) · unlock with TL Creds
  • Templated Extraction Burst and Reasoning Leakage in Output — C2/Collection (KQL) · unlock with TL Creds
  • Encrypted Reasoning Replay With Decrypt/Transcribe Prompt — Exploitation (Sigma) · unlock with TL Creds
  • Shared Prompt Across Many New LLM Accounts (Account-Farm Cluster) — Post-Exploitation (Sigma) · unlock with TL Creds
  • Templated Extraction Burst and Reasoning Leakage in Output — C2/Collection (Sigma) · unlock with TL Creds
  • Encrypted Reasoning Replay With Decrypt/Transcribe Prompt — Exploitation (SPL) · unlock with TL Creds
  • Templated Extraction Burst and Reasoning Leakage in Output — C2/Collection (SPL) · unlock with TL Creds

Indicators of compromise

As of 2026-10-04, Threadlinqs Intelligence tracks 4 indicators of compromise for TL-2026-2915.

4 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

OpenAI attributed a core cluster to individuals associated with Moonshot AI (developer of Kimi) but said it is unclear whether all operators were a single actor.

OpenAI publicly disclosed on September 30, 2026 that it had disrupted a coordinated model-distillation campaign targeting the protected (encrypted) reasoning of its models. Per OpenAI as relayed by secondary reporting, the operators did not break the encryption or access any database or customer conversation. Instead they manipulated model interactions: encrypted reasoning produced in one conversation was copied into a second conversation, where the model was instructed to decrypt and transcribe it, using attacker-controlled prompts and reusable context. OpenAI closed a pathway that let someone who already held another user's encrypted reasoning replay it and recover its contents.

Timeline and scale: activity began around July 1, 2026 at low volume, spiked on July 24-25 with roughly 16,000 requests matching the extraction pattern from more than 4,000 users, and a broader sweep for related prompt-pattern activity identified a cluster of more than 15,000 users. The whole cluster was fully disrupted by July 28, 2026.

Attribution: OpenAI attributed a core cluster of the activity to individuals associated with Moonshot AI, the developer of Kimi, while stating it is unclear whether all operators observed in the period originated from a single actor. Reporting notes OpenAI did not establish involvement of Moonshot founder Yang Zhilin or company leadership. No Moonshot response was documented in the reviewed articles.

Context: the technique class was independently described on August 10, 2026 in the arXiv paper 'Stealing Reasoning Traces from Proprietary LLM APIs', which reported that encrypted reasoning blocks from OpenAI, Anthropic and Google could be replayed into weaker sibling models to recover plaintext traces. Earlier, in February 2026, Anthropic reported industrial-scale distillation of Claude by DeepSeek, Moonshot and MiniMax via roughly 24,000 fraudulent accounts and over 16 million exchanges.

Response: OpenAI banned or restricted fraudulent accounts, tightened signup and infrastructure controls, expanded monitoring for related accounts, coordinated with third-party services the activity ran through, closed the replay pathway, added checks that detect and hold streamed output that might expose reasoning, strengthened cross-user and cross-organization protections, and shared intelligence via the Frontier Model Forum and government channels. No CVE, CVSS score, or network indicators (IPs, domains, hashes) have been published. The OpenAI primary report could not be fetched directly (HTTP 403) and facts here come from consistent secondary reporting.

MITRE ATT&CK / ATLAS techniques by tactic:

AI Attack Staging

AML.T0005 Create Proxy AI Model

Exfiltration

AML.T0024 Exfiltration via AI Inference API; AML.T0024.002 Extract AI Model

AI Model Access

AML.T0040 AI Model Inference API Access

Execution

AML.T0051 LLM Prompt Injection

Resource Development

T1585.003 Establish Accounts: Cloud Accounts

Remediation

Patches

  • OpenAI closed the encrypted-reasoning replay pathway server-side; no customer action required

Immediate actions

  • Block or restrict the fraudulent accounts identified in the cluster
  • Hold or inspect streamed model output that may expose protected reasoning
  • Monitor for related accounts and shared infrastructure

Workarounds

  • Do not publish raw agent transcripts that contain encrypted reasoning blocks

Longer-term hardening

  • Bind encrypted reasoning artifacts to the originating user, organization and conversation so they cannot be replayed cross-account
  • Strengthen signup and infrastructure controls against bulk fraudulent account creation
  • Share distillation-abuse intelligence through industry bodies such as the Frontier Model Forum

Affected and fixed versions

  • OpenAI — OpenAI model APIs (protected/encrypted reasoning)
    Vulnerable versions: Prior to replay-pathway fix, July 2026
    Fixed in: Replay pathway closed by OpenAI (server-side)

Timeline

  • Anthropic publicly reports industrial-scale distillation of Claude by DeepSeek, Moonshot and MiniMax via roughly 24,000 fraudulent accounts and over 16 million exchanges (precedent for the same class of abuse).
  • Low-volume reasoning-extraction activity against OpenAI models begins.
  • Spike begins: ~16,000 requests matching the extraction pattern over July 24-25 from more than 4,000 users.
  • Peak ends; a broader sweep links related prompt-pattern activity to a cluster of more than 15,000 users.
  • OpenAI fully disrupts the cluster, banning or restricting fraudulent accounts.
  • arXiv paper 'Stealing Reasoning Traces from Proprietary LLM APIs' shows encrypted reasoning blocks can be replayed into weaker sibling models to recover plaintext across OpenAI, Anthropic and Google.
  • OpenAI publishes its report attributing a core cluster to individuals associated with Moonshot AI, with caveats on single-actor attribution.
  • GBHackers reports the disruption and mitigations.

Sources cited

Previous and next in threat intel

Drafted by the Threadlinqs research pipeline from 8 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats