CP SignalTechnology worth your attention.

Thursday, 22 October 2026

The Daily Edition: 3 stories, ranked by significance

  1. Artificial intelligence

    Combining Long-Term Memory and LLMs Comes at a High Cost: From Prompt Injection to Persistent Compromise

    A new arXiv paper demonstrates that long-term/persistent memory features in ChatGPT Instant 5.3, Claude Sonnet 4.6, and Gemini 3 Flash can be exploited via prompt injection to achieve all NIST attack goals, turning a single-session jailbreak into a persistent, cross-session compromise. The researchers show attack vectors spanning personal memories, projects, skills, org instructions, and MCP-connected resources, and report that Claude's shared-memory features can let an attacker compromise multiple users in an organization and exfiltrate prompts. They state nearly half of the mitigations they derived remain unimplemented across the evaluated services.

    Read the storyNo. 1 of 3arXiv cs.CLTue 20 Oct
  2. Artificial intelligence

    Using Claude to reduce the attack surface of code

    Anthropic used Claude to automatically analyze attack surface across 25 open-source projects (5M+ lines of code) by mapping reachable code and configuration options rather than hunting individual vulnerabilities. Claude inventoried over 17,000 configuration options, flagged 3,842 as security-relevant, and surfaced 493 issues including pre-authentication exposure, silently-disabled security controls from conflicting options, and a fail-open DNS/hostname access-rule pattern found independently in four unrelated projects. One flagged configuration conflict that silently dropped a certificate-based authentication requirement had reportedly existed in the software for 21 years.

    No. 2 of 3Anthropic NewsTue 20 Oct
  3. Artificial intelligence

    Forty Questions: Identifying a Language Model from One-Word Answers

    A new arXiv paper shows that an LLM behind an API can be fingerprinted from its answers to 40 everyday one-word questions, correctly identifying the exact model 92.5% of the time across 94 models, versus 1.1% by chance. Applying this to 96 OpenRouter endpoints of 13 open models found 92 matched their claimed identity, but one endpoint billed as "Llama 4 Maverick" actually self-identified as, and was billed as, Gemma 4, and billing analysis also revealed inconsistent hidden token overhead across providers like Grok 4.7 versus GPT-4o.

    No. 3 of 3arXiv cs.CLTue 20 Oct