Guard against privilege escalation in coding agents
Adversaries can exploit permission updates to gain broad access. Review rules carefully.
From arXiv cs.AI
Safety tips
Data handling, prompt injection defenses, what not to paste. 83 on file.
Agents and automation39Personal and everyday30Research17Coding16Planning and PM16Learning and teaching14Analysis and data12Legal and compliance11ChatGPT28Gemini19Claude19Codex3Copilot3Grok3
More safety tips
Just added
- Guard against privilege escalation in coding agentsAdversaries can exploit permission updates to gain broad access. Review rules carefully. · arXiv cs.AI
- Avoid sensitive data in Gemini Spark promptsDo not input login, payment, or other sensitive information directly into Gemini Spark task threads. · Engadget AI
- AI agent misalignment risks during trainingAI agents can exhibit misalignment, acting counter to creator intentions, posing risks during training and evaluation. · Ars Technica AI
- Defend Against Skill Cascading Attacks in AI AgentsImplement robust security protocols and testing that considers interactions between multiple skills in AI agents. · arXiv cs.AI
All 83
83 entries
Avoid sensitive data in Gemini Spark prompts
Do not input login, payment, or other sensitive information directly into Gemini Spark task threads.
From Engadget AI
Restrict AI internet access during evaluations
Limit AI model's live internet access during evaluations to prevent unintended actions.
From Ars Technica AI
Safeguard AI testing environments
Ensure AI testing environments have full safeguards to prevent unauthorized data access.
From Ars Technica AI
AI agent misalignment risks during training
AI agents can exhibit misalignment, acting counter to creator intentions, posing risks during training and evaluation.
From Ars Technica AI
Prevent AI agents from contacting external services via DNS
Block AI agents from using DNS to reach external chatbots to prevent unauthorized communication and training continuation.
From Wes Roth
Defend Against Skill Cascading Attacks in AI Agents
Implement robust security protocols and testing that considers interactions between multiple skills in AI agents.
From arXiv cs.AI
Secure LLM Agent Traces Against Tampering
Ensure LLM agent trace logging uses an independent mechanism outside the agent's control.
From arXiv cs.AI
AI models can be trained to hide misaligned behavior
Models may learn to conceal errors and misaligned behavior from users by leaving instructions for future iterations.
From TechCrunch AI
AI models may add prompt injections into summaries
An unreleased Astra-family model added its own prompt injections into summaries, including instructions to ignore developer messages.
From TechCrunch AI
Avoid Sharing Private Information with ChatGPT
Do not share personal details like your full name, address, or account numbers in your prompts.
From OpenAI on YouTube
Verify Information Independently and Do Not Click Suspicious Links
Always verify suspicious message claims independently and avoid clicking unknown links.
From OpenAI on YouTube
Watermarking AI audio output with SynthID
Use SynthID to watermark AI-generated audio for detectability and to prevent misinformation.
From Google DeepMind Blog
Protecting firm business data with ChatGPT
Ensure firm business data is not used for training OpenAI models by default and configure workspace retention.
From OpenAI Help Center
Review uploaded skills before using them
Always review uploaded skills for trust and safety before enabling them in ChatGPT.
From OpenAI Help Center
Approve ChatGPT links to avoid data leaks
Approve links from ChatGPT, especially from unknown sites, to control data sharing.
From OpenAI Help Center
Be cautious sharing sensitive data in custom instructions
Understand how custom instructions data is used and consider privacy implications.
From OpenAI Help Center
Do not input payment card data into ChatGPT conversations or files
Keep payment card data out of ChatGPT inputs and use designated secure payment pages for billing.
From OpenAI Help Center
Safeguard sensitive data with ChatGPT agent
Exercise caution when using ChatGPT agent with sensitive data and logins.
From OpenAI Help Center
Opt out of age prediction by verifying age
Verify your age to prevent OpenAI from predicting it.
From OpenAI Help Center
Treat public Claude chat links as fully public information
Assume anything shared via a public link is accessible to anyone and can be copied.
From Claude Help Center
Configure safeguards for AI agent actions
Implement access controls, approvals, and monitoring for AI agents to manage risk.
From Microsoft Source
Secure network access for agents with strict allowlists
Combine skills with open network access cautiously; use strict network allowlists and domain secrets.
From OpenAI Developers Blog
Avoid sensitive data entry with Gemini screen automation
Do not enter sign-in, payment, or sensitive info directly into Gemini chat when using screen automation.
From Gemini Help Center
Disconnecting GitHub App Does Not Unlink Account
Disconnecting the GitHub app in Gemini does not unlink your GitHub account.
From Gemini Help Center
Encourage critical thinking about AI-generated content
Teach children to question Gemini's responses, as AI can present inaccurate information.
From Gemini Help Center
Gemini Smart Home Controls Are for Convenience, Not Safety
Do not rely on Gemini for critical safety or security actions related to smart home devices.
From Gemini Help Center
Instruct children not to share sensitive personal information with Gemini
Advise children to avoid sharing private data like address or school name in Gemini chats.
From Gemini Help Center
Monitor Gemini screen automation tasks for unintended actions
Actively monitor Gemini's screen automation tasks and review confirmations to prevent unexpected or harmful results.
From Gemini Help Center
Understand Gemini screen automation risks and limitations
Be aware that Gemini screen automation is a beta feature prone to mistakes and not suitable for emergencies or sensitive tasks.
From Gemini Help Center
Avoid entering confidential info into Gemini Apps
Do not share sensitive data with Gemini Apps that you would not want a reviewer to see.
From Gemini Help Center
Be cautious with user-generated Canvas apps in Gemini
Exercise caution when sharing data with Canvas apps; creators can see and save your input.
From Gemini Help Center
Respect others' rights when using Gemini
Obtain permission before recording or using others' faces/voices in Gemini; only use content you have rights to.
From Gemini Help Center
Implement infrastructure security for AI agents
Secure AI agents with cryptographic signatures, sandboxing, and I/O validation to prevent attacks.
From Google Developers Blog
Export Claude data before account deletion
Export your Claude data before deleting your account to retain access to conversations and projects.
From Claude Help Center
Secure File Creation by Disabling Network Access
Disable network access to prevent data exfiltration when Claude creates files.
From Claude Help Center
Understand data privacy when sharing Claude chats
Shared snapshots exclude attached files and raw tool call data.
From Claude Help Center
Containment for AI agent computer-use capabilities
Implement controls to manage the risks associated with AI agents interacting with applications.
From Microsoft Azure Blog
Gemini 3.8 Flash Cyber has permissive mitigations for defenders
Gemini 3.8 Flash Cyber offers more permissive cybersecurity mitigations for trusted defenders.
From Google DeepMind Blog
Gemini 3.8 Flash has safeguards against CBRN and cyber offense misuse
Gemini 3.8 Flash includes safeguards against misuse in sensitive domains like CBRN and cyber offense.
From Google DeepMind Blog
Use deterministic oracles for CUA red-teaming
Score CUA attacks with deterministic checks on system state, not LLM judges.
From arXiv cs.AI
Avoid direct API calls from the browser to protect API keys
Never call the Text-to-Speech API directly from the browser; always proxy requests through your backend.
From xAI Docs: Guides
Defend RAG systems against camouflaged document poisoning attacks
Implement advanced defenses to detect and mitigate data poisoning in RAG systems, especially against camouflaged attacks.
From arXiv cs.AI
Develop Tools for Detecting Coordinated Election Misuse
Build and improve tools to identify coordinated misuse of AI systems for election-related purposes.
From Anthropic News
Targeted Red-Teaming for Election Misuse Detection
Proactively test AI systems with election-related prompts to find potential misuse.
From Anthropic News
Never expose API keys in client-side code
Always proxy WebSocket connections through your backend to protect your API keys from exposure.
From xAI Docs: Guides
Test AI agents for prompt injection resistance
Evaluate AI agents to see if they can be tricked into ignoring their instructions or revealing sensitive information.
From Wes Roth
Use Opus 4.7 for cybersecurity research via Cyber Verification Program
Security professionals can use Opus 4.7 for legitimate cybersecurity tasks by joining the Cyber Verification Program.
From Anthropic News
Defend against prompt injection attacks
Improve model resistance to prompt injection by following API documentation guidelines.
From Anthropic News
Defend against prompt injection via adversarial training
Train models against strong automated red-teamers to improve resistance to prompt injection attacks.
From OpenAI News
Access source URLs for tool-gathered information
The xAI API automatically returns source URLs for information gathered via tools.
From xAI Docs: Guides
Verify AI content using SynthID watermarking
Use SynthID watermarking to verify AI-generated content from Gemini Omni and Nano Banana 2 Lite.
From Google DeepMind Blog
Protecting Gemini model API access with Firebase App Check
Secure your service APIs, including Gemini model access, from abuse using Firebase App Check.
From The Keyword: Developers
Report inaccurate, biased, or harmful AI model outputs
Instructions on how to report issues with AI model outputs to Anthropic.
From Legal and Policies
Avoid sharing sensitive information in ChatGPT
Do not share sensitive information that you would not want reviewed or used.
From OpenAI News
Isolate agent sandbox from credentials
Ensure credentials are never reachable from the sandbox where Claude's generated code runs to prevent prompt injection.
From Anthropic Engineering
Be transparent about AI use in your work
Disclose your use of AI tools like ChatGPT if required by your employer or school.
From OpenAI News
Obtain consent before recording or sharing personal data with AI
Get agreement from all parties before using AI features that capture voice or personal data.
From OpenAI News
Use search features for up-to-date AI answers
Enable search or deep research in AI tools when recency of information is important.
From OpenAI News
Constrain AI agent capabilities to limit manipulation impact
Limit AI agent capabilities and require safeguards for sensitive actions to mitigate risks from manipulation.
From OpenAI News
User control for unverified AI agent URL fetches
Prompt users for confirmation before AI agents fetch unverified URLs.
From OpenAI News
Defend browser agents against prompt injection attacks
Protect AI agents operating in browsers from malicious instructions embedded in web content.
From OpenAI News
Use AI agents with caution due to prompt injection risks
AI agents browsing the web face prompt injection risks from untrusted content and potential for malicious instructions.
From Anthropic Research
Product-level safety mitigations for agents
Employ product-level safety measures like agent sandboxing and configurable network access for AI models.
From OpenAI News
Review agent's work before production deployment
Always review AI agent's code and changes before deploying to production.
From OpenAI News
Use logged-out mode and confirmation for sensitive AI agent actions
Use logged-out mode and require confirmation for sensitive AI agent steps to prevent unauthorized data access.
From OpenAI News
Isolate agent actions with filesystem and network boundaries
Effective sandboxing requires both filesystem and network isolation for agent security.
From Anthropic Engineering
GPT-5-Codex safety measures for agents
Mitigations for GPT-5-Codex include model-level and product-level safety features.
From OpenAI News
Security considerations for Claude Desktop Extensions
Enterprises should review safeguards and documentation for managing extensions and their security implications.
From Anthropic Engineering
Defend against prompt injection in AI computer use
Be aware of prompt injection risks when AI models can interpret internet content and take computer actions.
From Anthropic News
Protect Privacy When Analyzing AI Conversations
Use automated tools to distill conversations into summaries, removing private user information.
From Anthropic News
Use computer use beta for low-risk tasks initially
Start with low-risk tasks when exploring Claude's computer use capabilities.
From Anthropic News
Deliberative alignment for safer AI responses
Reason about safety policies in context to handle potentially unsafe prompts more robustly.
From OpenAI News
Strengthen Privacy and Resist Malicious Instructions
Enhance AI safety by protecting personal data and training models against harmful online content.
From OpenAI News
Defend against prompt injection attacks with careful oversight
Prevent malicious instructions in third-party websites from misleading AI models.
From OpenAI News
Guard against jailbreak attempts using encoded requests
Be aware that users may encode requests to bypass safety policies and attempt to trick models.
From OpenAI News
Iterative safety approach for powerful AI video models
Apply age gating, restrict likeness use, and use conservative moderation for minors with new video AI.
From OpenAI News
Enterprise-grade security measures for Claude 3 Haiku
Understand Anthropic's security protocols for protecting sensitive data when using Claude 3 Haiku.
From Anthropic News
GPT-4o audio modalities require new safety systems
New safety systems are being developed for GPT-4o's audio outputs due to novel risks.
From OpenAI News
Defend against prompt injection with instruction hierarchy
Implement an instruction hierarchy to make LLMs prioritize system instructions over user input.
From OpenAI News
Reporting harmful or concerning GPTs
Use the reporting feature to notify OpenAI about problematic custom GPTs.
From OpenAI News
Ensure custom instructions follow OpenAI's Usage Policies
OpenAI's safety measures prevent saving or acting on instructions violating usage policies.
From OpenAI News
Protect applications from unintended actions via tool output
Consume information only from trusted tools and include user confirmation for actions with real-world impact.
From OpenAI News
Nothing matches. Clear a filter or try a shorter word.