Prompt injection: the agent itself is the vulnerability
„AI assistants and agents open a new, hard-to-close attack surface because any ingested content can become an instruction.”
Counter-narrative
Prompt injection is a known, fixable class of vulnerabilities. The sophistication of real-world abuse is so far low, and vendors have patched issues in several cases (for example, Microsoft fixed Copilot flaws). This position is represented by Google data and by the platform vendors that published fixes.
Voices
Noma Labs, OpenAI, Google, Microsoft, Dark Reading, AI security startups
Analysis
The narrative holds that the mixing of agent permissions and ingested content is a fundamental design flaw. A continuous stream of researcher disclosures fuels it, and almost every major platform is affected. AI security startups have an interest in creating a market, while platforms have an interest in framing it as an isolated, fixed bug. Its weak point is that most documented cases are researcher disclosures, and according to Google, the sophistication of real-world attacks is so far low.
What's at stake: The stakes are whether agent adoption in enterprises outpaces the development of authorization and isolation patterns.
Narrative sub-threads
Silent data leaks from corporate assistants
Corporate AI assistants that read email, calendars and documents can be induced to leak the data connected to them with a single crafted email, invitation or link, in several cases without any user click.
The actors are assistants from major platforms (Copilot, Gemini, ChatGPT, Rovo, Claude Desktop). The evidence is a series of zero-click and one-click vulnerabilities reported by research firms, some of which were patched by the manufacturer. In the case of Copilot Studio, data leaked even after the fix.
→ lasting · evidence: strongCoding agents as a supply-chain gap
AI coding agents and their MCP tools run code, disclose secrets and private data, or install malware in response to instructions hidden in repositories, comments, issues, configuration files and downloaded files.
The target is the developer workflow (Cursor, Claude Code, Gemini CLI, Copilot, Cline, Serena, MCP servers). The mechanism is an instruction planted in a code repository or file. The consequence may be supply chain compromise: in the Cline case, OpenClaw could have been installed on users between December 21, 2025, and February 9, 2026, due to a prompt injection flaw.
→ lasting · evidence: moderateBrowser agents on the open web
AI browsers and browser-embedded agents can be hijacked via the content or headers of visited websites, or through malicious extensions.
The attack surface is the public web and the browser extension, not the corporate data store or codebase. Evidence mainly consists of vulnerability disclosures and analyses regarding OpenClaw, Chrome Gemini Live, and the Claude extension. SquareX alleges the Comet vulnerability; Perplexity disputes it.
↗ rising · evidence: strongAgents that issue money and accounts
Agents authorized for transactions and administration have already caused real damage through manipulation: they emptied a crypto wallet linked to Grok, assisted in taking over Instagram accounts, fell for phishing, and disclosed user data. A prompt requesting authentication credentials also bypassed AWS AgentCore protection.
Here, damage is caused not by a leak but by executed actions. The evidence stems largely from live incidents rather than laboratory research: examples include the Grok wallet and a Meta AI-based customer service tool, through which, according to a Meta report, more than 20,000 Instagram accounts were compromised.
→ lasting · evidence: moderateManufacturer defense and admitted unpatchability
Manufacturers are increasingly integrating protections (Chrome hardening, Lockdown Mode, AI-based red teaming for GPT-5.6). However, OpenAI itself acknowledges that AI browsers may never be fully protected against prompt injection, and Check Point states that Lockdown Mode can be bypassed.
In this strand, model developers and measurement data are the key players. This includes defensive announcements and Google data indicating that malicious prompt injection attacks are increasing in number but remain low in sophistication. One report states that over 90 companies were hit by such attacks in 2025. This strand also includes the bypassing of Lockdown Mode.
Sub-thread stories
Silent data leaks from corporate assistants
Coding agents as a supply-chain gap
Browser agents on the open web
Agents that issue money and accounts
Manufacturer defense and admitted unpatchability
Who is affected and how
- security leader
- Before deploying Copilot, Rovo, and GitHub agents, request a data flow map, as one-click and zero-click data theft runs on corporate permissions.
- security operations
- Due to active exploitation of Flowise and Langflow instances, agent-building tools should be added to the vulnerability management inventory.
- developer
- Grant agent tool access based on the principle of least privilege. Treat external content (issues, emails, calendar invites) as untrusted input and restrict outgoing requests with allowlists.
- citizen
- Due to account theft via AI customer service, protect the account with strong, non-SMS-based two-factor authentication.
Hungary
No or only a minimal domestic trace in the examined period: among the stories, there is no item marked [HU].
Weekly trend
Turning points
Related narratives in the News magazine
Supporting stories
Machine analysis based on story titles and summaries, with a verification pass. The source of the facts is the article linked at each story.
What do you make of this piece?
The split shows after 20 votes.
Why? Your answer is not shown publicly; it helps the editors.