#ModelSecurity
AI agents attack LiteLLM, Dify & OneAPI via exposed creds—scale & automation turn minor flaws into major breaches. #AI #APIKeys #Cybersecurity #AIThreats #ModelSecurity #ARTEX https://thedailytechfeed.com/amazon-exposes-ai-agents-stealing-api-keys-from-litellm-dify-more/
November 16, 2026 at 11:11 AM
November 11, 2026 at 12:29 PM
November 11, 2026 at 7:30 AM
October 26, 2026 at 3:01 PM
Research exposes a vulnerability in TEE-based weight obfuscation for on-device LLMs, showing methods like GroupCover and ArrowCloak can be broken via the Solution Determinacy Property. The findings raise concerns about the…

#OnDeviceLLM #ModelSecurity #AIHardware
https://arxiv.org/abs/2610.21565
October 16, 2026 at 4:01 PM
October 15, 2026 at 12:14 PM
OpenAI disrupted Moonshot AI’s effort to extract hidden reasoning from its models via distillation attack. #AI #SecurityNews #AdversarialDistillation #MoonshotAI #OpenAI #ModelSecurity https://thedailytechfeed.com/openai-thwarts-reasoning-extraction-campaign-tied-to-moonshot-ai/
October 1, 2026 at 12:55 PM
September 30, 2026 at 10:51 AM
Researchers describe a steganography attack that hides medical images inside neural network weights via a StyleGAN2-based autoencoder, bypassing typical sanitization and statistical auditing on collaborative…

#MedicalAI #ModelSecurity #Steganography #AIResearch
https://arxiv.org/abs/2609.31726
September 29, 2026 at 8:01 PM
Repeated AI sandbox breakouts are raising fresh doubts about whether frontier AI labs can secure advanced models during testing, and what that means for deployment. #OpenAI #AISafety #ModelSecurity
AI Breaks Out Of Test Sandbox Again
OpenAI’s repeated AI sandbox breakout issues have raised concerns about whether frontier AI developers are mature enough to secure advanced models during testing, let alone in real-world deployment. The situation has intensified debate over how AI safety and security risks will be managed as these systems become more capable. #OpenAI
www.hendryadrian.com
September 29, 2026 at 12:00 PM
CISA flags China-based firms for extracting billions of tokens from models like GPT, Claude & Grok in large-scale mimicry attacks. #AI #Security #ModelSecurity #China #CISA https://thedailytechfeed.com/cisa-chinese-ai-firms-harvest-billions-of-tokens-from-us-frontier-models/
September 9, 2026 at 9:17 AM
August 28, 2026 at 6:49 PM
The Anthropic-Alibaba case will test whether "exporting" AI capability means shipping silicon or simply answering API calls. That answer will reshape how every AI company thinks about international access to their models. #AIPolicy #ModelSecurity #ExportControls
June 25, 2026 at 8:08 AM
OpenAI, Anthropic, and Google are sharing intelligence through Frontier Model Forum to combat Chinese AI model theft. Anthropic found 16 million unauthorized exchanges from three Chinese companies using 24,000 fraudulent accounts. #AI #ModelSecurity #Coordination
April 13, 2026 at 11:02 AM
Anthropic’s Claude Security enters public beta for repo vulnerability scanning with detailed repro steps. Cisco releases Model Provenance Kit for AI model tampering detection. AI-assisted phishing like Bluekit rises. #AIprivacy #ModelSecurity #USA
Cybersecurity News | Daily Recap [01 May 2026]
Daily Recap, AI Security updates highlight Claude Security's public beta for repository vulnerability scanning and Dataiku's Kiji Privacy Proxy to locally mask PII before prompts reach external AI APIs. The report also notes governance gaps with Shadow AI, Cisco's Model Provenance Kit for fingerprinting AI models and detecting tampering, and the emergence of AI-assisted phishing like Bluekit, along with other ransomware, supply-chain, and vulnerability news across Windows, SAP, and related ecosystems. #ClaudeSecurity #BluekitPhishing
www.hendryadrian.com
May 2, 2026 at 2:45 PM
Downloading Gemma 4 from Hugging Face involves risks like remote code execution via pickle-based formats and sleeper-agent backdoors in weights. Use safetensors, verify SHA-256 hashes, and check uploader identities. #ModelSecurity #DataPrivacy #USA
You Downloaded Gemma 4 from Hugging Face. Is It Safe to Run?
Local open-weight models like Gemma 4, Llama 4, and Qwen 3 preserve data privacy but introduce significant supply-chain risks when weights and serialization artifacts are downloaded from public hubs. Pickle-based formats enable remote code execution, model weights can contain sleeper-agent backdoors, and operators must require safetensors, hash verification, uploader vetting, and isolated testing to mitigate those threats. #Gemma4 #HuggingFace #Safetensors #Picklescan #Anthropic #CrowdStrike
www.hendryadrian.com
April 16, 2026 at 6:30 AM
Local AI models protect privacy but risk supply-chain attacks via pickled files and fine-tuned sleeper agents triggered by prompts. Use SafeTensors, verify hashes, and prefer trusted sources. #SafeTensors #ModelSecurity #OpenAI
Is Your Local AI Model Backdoored by Your Politics? Sleeper Agents Exposed
Local models preserve data privacy but introduce supply-chain security risks because downloaded model files (often pickled) can execute arbitrary code and fine-tuned weights can hide sleeper agents that trigger on specific prompts. Mitigations are simple and effective: download from verified providers on Hugging Face, prefer SafeTensors format, and verify model hashes to eliminate the vast majority of threats. #Pickle #SafeTensors #HuggingFace #DeepSeekR1 #PyTorch
www.hendryadrian.com
April 13, 2026 at 11:15 AM
Guide maps RAG and LLM risks (prompt injection, data/model poisoning), details baseline controls across data, model, and infrastructure layers, and offers high-risk model considerations. #AI #ModelSecurity #RAG https://bit.ly/3ZxghAO
February 13, 2026 at 8:03 PM
For years now the largest data centers leaned on an arrangement most people never saw itemized.

https://sharedsapience.com/century-report/the-century-report-august-7-2026/?ref=bs
#Modelsecurity #Aialignment
August 8, 2026 at 3:22 PM
DeepMind's Cyclone Forecaster Adds Track Warning Against ECMWF in Two Basins - and Google Hands It to the Forecasters

https://sharedsapience.com/century-report/the-century-report-august-7-2026/?ref=bs
#Modelsecurity #Aialignment
August 7, 2026 at 8:26 PM
The Other Side of the Future

For years, the cheapest way to fund the data-center buildout was to keep its cost where no investor could see it.

https://sharedsapience.com/century-report/the-century-report-july-22-2026/
#Modelsecurity #Agenticai
July 23, 2026 at 3:56 PM
The Hidden Debt Behind the Buildout Reaches $1.65 Trillion

The Century Report covered the roughly $350 billion in combined big-tech debt on July 13.

https://sharedsapience.com/century-report/the-century-report-july-22-2026/
#Modelsecurity #Agenticai
July 22, 2026 at 10:46 PM
Turns out some AI models are just remixing old papers and passing them off as new. A security researcher uncovers the sneaky re‑packaging trick—what does this mean for benchmarks and trust? Dive in to see the details. #AIPlagiarism #ModelSecurity #ResearchIntegrity

🔗 aidailypost.com/news/securit...
August 5, 2026 at 9:53 PM
August 21, 2026 at 11:39 PM
A vendor audit found unpinned models and trust_remote_code=True silently running unreviewed code, revealing the AI era's overlooked supply chain risk. #modelsecurity
Nobody Reviewed the Model. They Just Reviewed the Code Around It
hackernoon.com
July 1, 2026 at 9:00 AM