How to Interact with AI Models Without Exposing Sensitive Data
Conversations with public AI models such as ChatGPT, Gemini, Claude and GigaChat are scanned by automated filters and may be reviewed manually by employees or contractors of the model providers. Corporate administrators and, in some cases, law-enforcement agencies can also access chat histories.
To reduce future exposure, users should first disable the option that allows the model to learn from their dialogues. In OpenAI settings this is done by turning off “Improve the model for everyone.” Google Gemini requires disabling chat history so that conversations are automatically deleted after 72 hours. Similar toggles exist for DeepSeek and Anthropic Claude, while Sber GigaChat offers no such control in its consumer interface.
Additional hygiene measures include replacing real names, phone numbers and other identifiers with placeholders such as {name} or {phone}. When large volumes of data must be processed, tools like Redacto or Гарда Маскирование can automate masking, although results still require manual verification.
Long-running chats accumulate contextual information; therefore old conversations should be deleted and new tasks started in fresh sessions. Custom GPTs or Gemini Gems can store persistent instructions so that behavior does not need to be re-explained each time.
Models downloaded from Hugging Face should be scanned for malicious payloads with utilities such as HiddenLayer Model Scanner. Prompt-injection risks can be mitigated by pasting any third-party prompt into a word processor and applying a uniform text color to reveal hidden commands.
The most private approach is to run models locally with Ollama. After installation, a small model such as llama3.2:1b can be pulled and executed entirely on the user’s hardware. Larger production models like qwen3-coder:30b require at least 40 GB of RAM. Running the service inside a Docker container further isolates it from the host system.
For hybrid use, the open-source client ChatBox can connect to both local Ollama instances and paid API endpoints such as Cloud.ru Evolution Foundation Models. Users paste their API key, select compatible models, configure embedding and reranker components for RAG, and upload documents to a knowledge base that the model consults during generation.
Related articles
Raft Develops Multilabel Guardrail Classifier Detecting 15 Risk Categories with 3x Speed and Cost Gains
Raft has released details on a custom multilabel guardrail classifier designed to scan both incoming prompts and model outputs for 15 distinct risk categories in real time. The system handles Russian and English text while maintaining independent thresholds for each category to balance false positives against critical misses. By switching from a PyTorch baseline to TensorRT inference on NVIDIA RTX 3090 hardware, the team achieved a 2.95x reduction in single-request latency and lowered inference cost to $0.062 per million requests. The architecture uses per-category expert query tokens plus a lightweight interaction transformer to capture correlations such as those between armament and violent content. Training relied on asymmetric loss functions and post-epoch per-class threshold tuning rather than standard binary cross-entropy. Benchmarking against nine open guardrail and toxicity models showed superior macro-F1 on rare but high-impact categories while remaining an order of magnitude cheaper than external LLM judges.
Zhou Hongyi Warns AI Tools Are Industrializing Vulnerability Discovery
At the Fourth Cyberspace Security Forum in Tianjin, 360 founder Zhou Hongyi stated that vulnerability mining is shifting from artisanal workshops to automated production lines, compressing discovery cycles from months or years down to hours. AI tools such as Mythos are standardizing and automating the process, enabling attackers to replicate elite hacker expertise at scale through distilled models and agent swarms. 360's own Tulongfeng platform has already discovered over 10,000 vulnerabilities since its June release, including long-hidden high-risk flaws in Windows, Office, OpenClaw, Flowise, and Codex. The emergence of multi-agent systems introduces new attack surfaces because compromised agents can autonomously collaborate and move laterally faster than human operators. Zhou described this as the "second one-way transparency," where offensive tradecraft becomes copy-pasteable via prompts and toolchains. Defenders are advised to adopt "model-versus-model" strategies, automate vulnerability intelligence workflows with SOAR, enforce strict agent permission audits, and integrate AI into their own code review and detection engineering processes.
Anthropic Fable 5.1 System Prompt Fully Leaked Hours After Launch Exposing 275000 Characters of Rules
Anthropic released its flagship Fable 5.1 model alongside Mythos 5.1 on September 2, achieving strong benchmark scores including 90 percent on ARC-AGI-2. Within hours, researcher Pliny the Liberator published the complete 275000-character system prompt on GitHub, far exceeding the company's official 27000-word disclosure. The leaked document details 46 built-in tools, strict copyright restrictions, memory classification boundaries, and behavioral constraints that function as an internal employee handbook. The incident highlights that model weights remain the true core while prompt-based guardrails create an attack surface once mapped. It also reveals privacy rules that permanently exclude storage of minor identities, criminal records, and self-harm indicators even when users disclose them. The leak underscores the growing gap between vendor transparency claims and actual runtime instructions governing frontier AI systems.
Fine-Tuning Prompt Injection Detector: Five Rounds, Four Failures and Gates Against Regression
AGIMA technical director Andrey Nepryakhin details the challenges of adapting an open-source GLiNER-based prompt injection detector for Russian corporate traffic. The system processes nearly one million requests monthly from eighty employees using agentic tools. Five fine-tuning rounds using LoRA adapters on microsoft/mdeberta-v3-base were conducted, yet only the second round reached production. Failures stemmed from regression on previously fixed attacks, imbalanced negative samples, and reliance on rule-based signals instead of the model. The team introduced cumulative corpora, person-based splits, and ablation checks against the base model to prevent silent rollbacks. A gradient mass rule limits any single phenomenon to 10-20 percent of training data to avoid precision or recall collapse.