Prompt injection: analysis of 10 November 2026

Back to the Chronology · Current news in this section

Photo from Open dataset of prompt-injection CVEs with a monthly threat index, Hacker News

Analysis

Prompt injection has moved from a research curiosity to real attacks on AI agents, and the security industry is spending heavily to defend against it.

Over the past month, reporting shows prompt injection and jailbreaks causing real-world harm: a jailbroken AI agent reportedly ran a two-week attack on a European telecom, a Codex jailbreak toolkit was found hiding adware, and a Brazilian court fined lawyers for hiding a prompt injection in a filing. At the same time, the defence market is consolidating, with CrowdStrike and Palo Alto Networks paying hundreds of millions for AI agent security startups. Researchers are finding injections in new places, from HTTP headers to log files and PDFs, while a steady stream of papers argues that classifiers alone cannot stop the attacks and that agents need system-level controls.

Key developments

  • Attacks in the wild. Malbear Labs found a jailbroken DeepSeek agent that spent two weeks attacking a European telecom, and researchers disclosed an OpenAI Codex jailbreak bundle hiding adware that modifies an AI agent's behaviour. These are among the clearest signs that attackers now weaponise agents rather than just chatbots. 19
  • Courts meet prompt injection. A Brazilian court sanctioned lawyers who hid a prompt injection in a filing to manipulate the court's AI, with a $16,000 fine, an early legal precedent for this kind of manipulation. 1213
  • Big money for agent security. CrowdStrike is buying Fiddler AI and Dreadnode for about $495 million to add agent observability and automated AI security testing, Palo Alto Networks is in talks to buy Clutch for $500 million, and OpenAI has engaged security startup Hacktron to test its systems. 101187
  • New injection surfaces. A study of 217 million web responses found prompt injections hidden in HTTP headers; red-teamers report injections through log files and PDFs; and a dataset now tracks prompt-injection CVEs with a monthly threat index. 232554
  • Research: agents as untrusted components. New papers scale attacks with agent swarms, measure how web-agent harnesses affect resilience, and propose deterministic policy enforcement, taint tracking and least-privilege auditing of tool calls, treating the model itself as untrusted. 2928303336
  • Standards and practitioners. OWASP lists prompt injection as the top risk for LLM applications, and practitioners at the [un]prompted conference and in blog posts compare defences, from positional scanners to judge models. 2121718

Themes

  • Filters are not a boundary. Several practitioners and papers argue that prompt-injection classifiers make poor security boundaries, pushing designs towards permissions, approvals and isolation around the model. 243031
  • Agents raise the stakes. Attacks now target agents that act, from telecom intrusions to coding agents controlling robot arms, so successful injections have real-world effects. 13239

What to watch

  • Further disclosures of agent-driven intrusions like the telecom case, and how vendors respond. 1
  • Whether the CrowdStrike and Palo Alto deals trigger more acquisitions of AI security startups. 108
  • Courts and regulators setting rules on manipulating AI systems used in legal and official processes. 12
Stories this analysis is based on (25)
  1. Jailbroken AI Agent Ran Two-Week Attack on European Telecom Infosecurity Magazine
  2. AI Security Schneier on Security
  3. Open dataset of prompt-injection CVEs with a monthly threat index Hacker News
  4. Hiding prompt injections in PDFs, and six judges that catch them Hacker News
  5. OpenAI taps AI security startup Hacktron to test systems Tech in Asia
  6. Palo Alto Networks in talks to buy AI agent security startup Clutch for $500M TechCrunch
  7. Researchers Uncover Prompt Adware Hidden in OpenAI Codex Jailbreak Toolkit The Hacker News
  8. CrowdStrike to shell out nearly $500M for 2 AI security startups as the market for securing agents takes off TechCrunch
  9. CrowdStrike Reportedly Paying $495 Million for AI Security Firms Fiddler, Dreadnode SecurityWeek
  10. Brazilian lawyers fined $16K for hiding a prompt injection in a court filing Hacker News
  11. Brazilian Court Sanctions Lawyers for Prompt Injection in Court Filing Hacker News
  12. Prompt injection scanners should be positional, not phrasal Hacker News
  13. Decision model as prompt-injection judge: 40 of 40 on 78 texts Hacker News
  14. LLM01: 2025 Prompt Injection Hacker News
  15. Prompt injection in HTTP headers: what 217M web responses show Hacker News
  16. Prompt injection classifiers are terrible security boundaries Hacker News
  17. Prompt injection through log files: what red-teaming our assistant found Hacker News
  18. Harness Matters: Measuring and Dissecting Security Resilience in Web Agents arXiv (research papers)
  19. Scaling Adaptive Prompt Injection Attacks with Agent Swarms arXiv (research papers)
  20. Treating the Model as an Untrusted Component: Deterministic Policy Enforcement, Human Approval and Taint Tracking Against Prompt Injection in Tool-Using Agents, with Validation on Open-Weights Models arXiv (research papers)
  21. On the Relative Security of Prompt Injection: Mechanised Definitions, Defences and Verdicts for LLM Agents arXiv (research papers)
  22. When Coding Agents Get Arms: Safety Risks Across the Digital-Physical Boundary arXiv (research papers)
  23. Least-Privilege Auditing of Agent Tool Calls with Intent Graphs and Reasoning-Guided Escalation arXiv (research papers)
  24. Agents of Chaos: A Systematization of Knowledge (SoK) on Malicious Attacks on Agentic AI Systems arXiv (research papers)
  25. Notwithstanding Anything to the Contrary, Ignore All Instructions: Measuring Prompt Injection Guards on Legal Contract Review Attacks arXiv (research papers)

Written by Claude (Anthropic), based on 40 recent headlines and excerpts in this section, Nov 10. AI analysis can misread or miss things: the numbers link to the original reporting, which is the authority.