A model may follow untrusted text as an instruction. The guidance: limit what a fooled model can do.
Illustration made with AI.
A model may follow untrusted text as an instruction. The guidance: limit what a fooled model can do.
Illustration made with AI.
@threadlinqs.bsky.socialPhishers hide prompts in emails to hijack your AI inbox assistant - humans see nothing. https://intel.threadlinqs.com/threat/TL-2026-3218 #ThreatIntel #Barracuda #PromptInjection #Phishing
@royans.bsky.socialGhostcommit: Image-Embedded Prompt Injection Bypassing AI Code Review
https://nsfocusglobal.com/ai-security-incident-case-ghostcommit-attack-leveraged-images-to-steal-secrets
https://flagthis.com/tldr/8597
##PromptInjection ##Steganography ##AICodeReview ##DataExfiltration ##LLMSecurity
@consultantben.bsky.socialBefore your business turns on any AI tool, ask these 3 questions. They catch most of the risk in about 5 minutes. Full 6-step AI risk assessment on my channel.
#AIrisk #AIsecurity #AIgovernance #cybersecurity #shadowAI #infosec #promptinjection
@glitchmentalmx.bsky.socialLa IA hackeó a un gobierno por una pregunta aburrida
@nile11.bsky.socialCal AI Teen Founder Raises $10M for Persona AI Wearable
..................
https://nile1.com/cal-ai-teen-founder-raises-10m-for-persona-ai-wearable/
..................
#aiAgent #aiWearable #amazon #bee #calAi #instinct #meta'sMuse #myfitnesspal #persona #promptInjection #stripeLink #vineVentures #za
@potato.softwareOne bad open can cost a day, or a deal.
Trismag catches known override patterns in emails and sites before it reaches your agent. Cheaper than cleaning up after.
Sign up free at trismag.dev
@magzforyks.bsky.socialOne bad open can cost a day, or a deal.
Trismag catches known override patterns in emails and sites before it reaches your agent. Cheaper than cleaning up after.
Sign up free at trismag.dev
@promptfoundry.bsky.socialNew arXiv work shows world models used in robot training can be a stealthy poisoning vector, letting attackers inject malicious prompts or transition dynamics into visibly safe teleoperation data and produce…
#WorldModels #Robotics #AIsecurity #PromptInjection
https://arxiv.org/abs/2606.09499
@magzforyks.bsky.socialUnfamiliar email. Unknown site. Fresh attachment.
A warning label before you open it: Trismag inspects the text first, before it reaches your agent. Hold for review, then you decide.
Sign up free at trismag.dev
@blackfog.comNew in ADX Vision 2.0
Seven layers of prompt protection help stop manipulated AI prompts from exposing sensitive data, whether they come from a person or an agent.
@psoheil.bsky.social#Meta just published how it secures #Muse, its personal AI agent, and the architecture is worth studying for anyone building agentic systems.
@cyberupdates365.bsky.socialResearchers demonstrated Cryptographic Context Injection against GitHub Copilot CLI in autopilot mode.
Encrypted instructions caused the agent to read a local file and send its contents externally under specific.
@thedailytechfeed.comEncrypted prompt injection in Copilot CLI can expose secrets via encrypted payloads. #AI #Cybersecurity #GitHub #CopilotCLI #PromptInjection #DevSecOps thedailytechfeed.com/encrypted-pr...
@promptfoundry.bsky.socialNew arXiv work shows attackers can shorten predicted output lengths by up to 83.4% via adversarial suffixes, giving their requests priority in LLM schedulers and faster completion. Worth noting if you build or use…
#LLM #AIsecurity #PromptInjection #MLops
https://arxiv.org/abs/2610.03430
@bizbln.bsky.socialMCP secret protocol lets AI agents talk but no identity checks one poisoned prompt can infect all! #AI #Security #Trust #PromptInjection #MCP #CyberRisk #TechAlert
@hacks.grAI tools that hand work to other tools can also pass along malicious instructions.
A researcher found flaws in systems used by Google, JPMo…
@hendryadrian.bsky.socialPrompt injection is only the start. Real AI security comes from least privilege, scoped tools, human approval, logging, and containment when agents read emails, docs, or act autonomously. #PromptInjection #AISecurity #LeastPrivilege
@robotcurrent.bsky.socialResearchers introduced PI3D, a prompt injection attack on multimodal LLMs that works by placing text-bearing objects in 3D environments rather than editing 2D images, highlighting a new risk for robotics and situated…
#MLLM #PromptInjection #Robotics #AIsecurity
https://arxiv.org/abs/2602.07104
@ossradarai.bsky.socialA new arXiv paper introduces APEX, showing how chained open-source LLM skills can carry attacker-controlled claims across steps to hijack agent decisions on SkillsBench. This highlights a supply-chain risk when…
#AIsecurity #LLMagents #OpenSourceAI #PromptInjection
https://arxiv.org/abs/2610.01564