Okta’s threat‑intelligence team says a 7 GB infostealer dump released on a Telegram channel on August 2, 2026 contained 1,843 unexpired JSON Web Tokens (JWTs) and encrypted JWTs (JWEs) that can be replayed to bypass multi‑factor authentication (MFA) and grant attackers direct access to premium large‑language‑model (LLM) APIs.
Scope of the dump
The dump covered data from 5,871 infected machines spread across 162 countries, according to the analysis published by The Hacker News on September 9, 2026. In total the dataset held 44,791 unique JWTs, of which 555 were identified as likely related to AI services such as OpenAI, Google Gemini and Anthropic.
Okta identified 2,937 JWE structures representing encrypted JWTs. Among all tokens, 17.7 % contained plaintext personally identifiable information (PII), raising the risk of credential‑theft beyond simple API abuse.
How replayable tokens bypass MFA
"Session tokens and API keys are sought specifically by threat actors because it is often possible to replay those secrets and bypass credential‑based authentication," said Jeremy Kirk, director of threat intelligence at Okta, in the report shared with The Hacker News. "Once successfully replayed, a threat actor is effectively logged in to an LLM service without actually logging in. Use of these skeleton keys makes abuse more challenging but not impossible to detect."
A valid JWT can be presented to an AI provider’s authentication endpoint in place of a fresh login, allowing the bearer to obtain a fresh access token and consume the service without satisfying password or MFA checks. Because the tokens were unexpired on the day of release, they could be used immediately.
Impact on leading AI providers
The analysis notes that most of the AI‑related tokens were set by OpenAI, which uses the NextAuth.js framework. The dump also included 24 still‑valid API keys for four AI‑related services, specifically Google Gemini, OpenAI, Groq and OpenRouter. While the packet does not break down the monetary value of the free usage, the ability to bypass MFA removes a key defensive layer for enterprise customers of these platforms.
Google, Microsoft and Anthropic – the three major AI players referenced in the source excerpts – each run multi‑billion‑dollar operations. Their chief executives are Sundar Pichai (Google), Satya Nadella (Microsoft) and Dario Amodei (Anthropic). All three companies are headquartered in the United States and employ tens of thousands of staff, underscoring the scale of the potential exposure.
Timeline and open questions
- 2026‑08‑02 – 7 GB infostealer dump released on a public Telegram channel.
- 2026‑09‑09 – Okta’s analysis published via The Hacker News.
The packet does not disclose how many of the 1,843 replayable tokens were sold on underground forums, nor does it provide a breakdown of which specific AI services each token would unlock. Okta also flagged 24 still‑valid API keys, but the total number of keys that may have been compromised remains unknown.
| Metric | Count |
|---|---|
| Total JWTs | 44,791 |
| AI‑related JWTs | 555 |
| Unexpired JWTs & JWEs | 1,843 |
| JWE structures | 2,937 |
| Infected machines | 5,871 |
| Countries represented | 162 |
Security teams are advised to rotate any API keys or tokens that may overlap with the dates in the dump and to monitor for anomalous usage patterns on LLM services. The incident illustrates a growing underground market for stolen AI credentials as enterprises increase AI adoption.
What remains unknown is the extent to which threat actors have already exploited the replayable tokens and the financial impact on AI providers. Okta’s report does not quantify the number of successful MFA bypasses, leaving a gap that further investigation will need to fill.