💥 BREACHES & INCIDENTS
- MonsterCloud CEO Zohar Pinhasi indicted for secretly paying ransoms while defrauding ransomware victims
MonsterCloud CEO Zohar Pinhasi was indicted on wire fraud charges for secretly paying ransoms while claiming to decrypt ransomware without paying hackers.- Applies to MonsterCloud clients, including local law enforcement agencies, businesses, and other ransomware victims who paid for data recovery services.
- Pinhasi allegedly claimed to use proprietary decryption tools while actually paying cybercriminals for decryption keys without clients' knowledge.
- He charged clients substantially more than the ransom paid — e.g., an $8,200 ransom billed at $150,000 — totaling over $19 million charged against more than $8 million in ransoms.
- Charged with two counts of wire fraud and one count of wire fraud conspiracy; if convicted he faces up to 20 years per count.
- The indictment mirrors ProPublica's 2019 'The Trade Secret' investigation into MonsterCloud's practices.
🕵️ RESEARCH & DEEP DIVES
Hirundo finds geopolitical bias in Alibaba's Qwen3-Coder weights, producing weaker code for US-aligned customers
Hirundo researchers showed Alibaba's Qwen3-Coder writes more vulnerable code when prompted for US-aligned customers, and that removing its CCP alignment from the weights eliminates the bias.- Applies to users of Alibaba's Qwen3-Coder-30B-A3B-Instruct and other Chinese open-weight models, which power Western products like Airbnb, Cursor, DoorDash, and Perplexity
- Asking for an app 'for a US defense contractor' instead of 'for a Chinese defense contractor' made the app 19% more likely to be hackable (70% vs 59%); apps for a Hong Kong democracy group were 28% more likely to leak users' secrets
- The bias lives in the model's weights — security prompts did not close the gap; with a security instruction, US contractor apps were 27% more likely to be hackable than a neutral customer's
- Testing used BaxBench (ETH Zurich/LogicStar), running thousands of app builds with only the customer name changed and attacking each running app with exploits
- Hirundo's 'Westernization' (behavioral unlearning removing CCP alignment) eliminated all three gaps without degrading coding performance; the fixed model is on Hugging Face
NIST, CISA, NSA publish guide on incident response for runaway AI agents, with identity records as key evidence
NIST, CISA, and NSA jointly published ten voluntary practices for incident readiness and response when AI agents act in undirected ways.- The guide 'Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation' was published October 16, 2026 by NIST, CISA, and NSA.
- It targets organizations training or evaluating highly capable agents on long-horizon tasks with limited human supervision, and defines a 'control incident' as agents taking significant undirected actions that undermine controls or oversight.
- The ten practices span five stages: prepare, detect, respond, recover, and report, building on established incident-management frameworks like least privilege, credential management, and evidence preservation.
- The guide treats prompt-injection or backdoored-model actions as 'instructed' and thus outside its control-incident definition, complicating responder triage where intent is unanswerable.
- The OpenCode breach (1.4% of Console users, exposed API keys for Anthropic, OpenAI, and Google, unauthorized charges up to $1,700) illustrates that agents accumulate credentials the organization holds but doesn't issue, and revocation presupposes an inventory of what the agent could touch.
Chinese national extradited from Netherlands for role in $36.9M crypto investment scam
A Chinese national was extradited from the Netherlands and arraigned in California for allegedly laundering $36.9 million stolen from 174 U.S. victims in a cryptocurrency investment scam.- Zheng Zhao, 44, a Chinese national, was arraigned in the Central District of California on conspiracy to commit money laundering charges and faces up to 20 years in prison.
- The scam targeted 174 U.S. victims, who lost over $36.9 million through fake cryptocurrency trading platforms promoted via unsolicited texts, social media, phone calls, and dating sites.
- Victim funds were funneled to a Deltec Bank account in the Bahamas under the name Axis Digital Limited, converted to Tether (USDT), and moved to wallets controlled in Cambodia.
- Eight co-conspirators have pleaded guilty, including Jingliang Su, sentenced to 46 months, and ShengSheng He, sentenced to 51 months.
- The case is part of the Justice Department's Cyber-Enabled Scam Initiative, investigated by the Secret Service's Global Investigative Operations Center.
Siemens, Schneider Electric, ABB Confirm Some Products Affected by PHP2Shell PHP Flaws
Major industrial vendors have published advisories confirming some of their products are affected by the PHP2Shell PHP vulnerabilities.- Siemens, Schneider Electric, ABB, Rockwell Automation and Johnson Controls published advisories on the PHP2Shell flaws (CVE-2026-109618 and CVE-2026-109619), critical PHP engine bugs allowing unauthenticated RCE.
- Affected products include Schneider Electric's EcoStruxure Security Expert, ProLeiT LINEoptimize and EcoStruxure Control Engineering; Siemens Ruggedcom APE1808 with PAN-OS 10.2/11.1/11.2/12.1/12.2; ABB MES deliveries with DokuWiki; and B&R APROL up to R 4.4-65.
- Belden's NetModule NRSW (all versions) and macmon 6.x/7.x are affected; Johnson Controls patched Illustra Flex4/Pro4/Gen5/Essentials Gen4 and Visonic PowerManage 5.0.45; HMS patched Netbiter firmware.
- PHP2Shell flaws stem from memory-safety bugs in PHP's error handling, affecting PHP versions back to 5.4; Wiz Research used frontier AI to develop working exploits, and PHP released fixes on October 21 (8.5.12, 8.4.27, 8.3.36, 8.2.35).
- First exploitation attempt was seen October 17 before patches; Cloudflare reported millions of attempts, and the first confirmed attacks hit honeypots over the weekend.
Genesys Hack Exposes Customer Data From Internal LMS and CRM Systems
Genesys says hackers accessed and likely downloaded customer data from its internal LMS and CRM systems.- Genesys, a customer experience and call center provider serving 8,000+ organizations, disclosed unauthorized access to internal business systems on October 19, 2026, after becoming aware of the incident on October 16.
- Data accessed and/or downloaded includes names, business email addresses, phone numbers, job titles, company and account information, and other customer interaction details.
- The Genesys Cloud, Genesys Engage, and Genesys Engage Cloud platforms were not compromised, and customer/partner data including call recordings and analytics data was unaffected.
- Genesys suspended the LMS platform and announced the incident was fully contained on October 23; the investigation into scope and impact is ongoing.
- A group called ScarlettGroup has claimed the attack on its leak site; Genesys warns customers to watch for suspicious communications impersonating Genesys.
Compaction Smuggling Flaw Let Researchers Steal OpenAI Model Reasoning for Cheap Distillation
Xint discovered a flaw in OpenAI's context compaction process that allowed reasoning traces to be moved between models, enabling cheap distillation of rival AI.- Affects OpenAI frontier models exposed via public APIs, including gpt-5.6-sol, gpt-4.1, and gpt-4o; follows an August 2026 paper showing encrypted reasoning could be replayed into a weaker model and transcribed.
- The 'Compaction Smuggling' attack bypassed OpenAI's August mitigations by encoding one model's reasoning inside a compaction summary, then asking a second model to output its input exactly as-is.
- The attack worked at every reasoning effort tested and enables large-scale distillation — using stolen reasoning to cheaply build rival models.
- OpenAI confirmed the issue, fixed it, and awarded Xint a $6,000 bounty on October 23, 2026.
- CISA warned in September that Chinese firms including DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI extracted billions of tokens from US frontier models since at least late 2024.