All CosmicBytez Labs articles tagged #Prompt Injection, across news, security advisories, how-to guides, and projects.
Researchers from the ASSET Research Group disclosed GhostSplice — a novel cross-channel trust fragmentation attack that splits malicious instructions across multiple MCP tool calls, bypassing safety filters to make AI coding agents like Cursor exfiltrate SSH keys, .env files, and source code.
As enterprise AI agent deployments accelerate, a dangerous pattern is emerging: agents assigned loosely defined tasks are being granted broad, persistent access to organizational systems — creating an attack surface that traditional identity and access controls were never designed to handle.
Researchers at Pillar Security documented the first confirmed real-world agent-to-agent exploitation in a production system, targeting Google's Agent Development Kit for Python — downloaded over 90 million times. A prompt injection via a malicious pull request could hijack a high-privilege maintainer AI agent and compromise the CI/CD pipeline.
Two independent security firms found that Atlassian's AI assistant Rovo is vulnerable to indirect prompt injection attacks that silently exfiltrate Jira and Confluence data to attacker-controlled servers — with no visible trace in the chat log.
Researchers discovered a parameter-to-prompt injection flaw in Atlassian Rovo AI that allowed a single malicious link click to exfiltrate Confluence API keys, Jira data, and files from Microsoft 365 and Google Workspace integrations.
A new class of prompt injection embedded in "Ask AI" buttons on commercial websites silently writes biased recommendations into users' AI assistant memory — no malware, no credentials, just a weaponized URL parameter.
As AI agents, copilots, and LLM-powered applications proliferate across enterprise environments, traditional packet-inspection firewalls are blind to the threats they introduce. Check Point argues the network layer — the universal chokepoint for all AI traffic — must evolve into an intent-aware enforcement platform.
A CVSS 9.9 critical vulnerability in PraisonAI before 4.6.78 allows attackers to write files to arbitrary filesystem locations and execute arbitrary OS...
A CVSS 10.0 critical vulnerability in PraisonAI before 1.6.78 allows attackers to achieve remote code execution by injecting malicious prompts that...
Security researchers have demonstrated 'Ghostcommit,' a technique that embeds prompt injection payloads inside PNG images to bypass AI code review tools...
Researchers at the AI Now Institute have demonstrated a 'Friendly Fire' attack that tricks AI coding agents — including Claude Code, Gemini CLI, and...
Two critical vulnerabilities (CVE-2026-50548, CVE-2026-50549) dubbed DuneSlide allow zero-click prompt injection attacks to escape Cursor IDE's sandbox...
Security researchers have found that classic Bash shell techniques — some dating back decades — can bypass the safeguards in most open-source AI coding...
OpenAI has begun rolling out a new Lockdown Mode to ChatGPT for eligible personal accounts, restricting tool capabilities that could be exploited in prompt…
Prompt injection vulnerabilities in Salesforce Agentforce and Microsoft Copilot would have allowed unauthenticated attackers to exfiltrate sensitive CRM...
A critical flaw in AI Code's automatic terminal command execution design allows unsafe commands to bypass the model-based safety judgement and be...
China's CNCERT has warned that OpenClaw (formerly Clawdbot/Moltbot), the viral self-hosted AI agent, carries over 250 disclosed vulnerabilities including...
Microsoft's Defender team tracked over 50 unique prompt injection payloads from 31 companies using 'Summarize with AI' buttons to manipulate chatbot...