Tag: Prompt injection

Aktualności12 sierpnia 2026

Auto Mode Becomes Default in Claude Code for Pro, Max, Team

Anthropic is making auto mode the default in Claude Code for the Pro, Max, and Team plans from August 14, 2026. Instead of asking for approval on every action, a classifier lets tool calls through, blocks them, or looks for a safer alternative.

Aktualności1 sierpnia 2026

A fundamental LLM flaw: forged chain-of-thought bypasses safety

Researchers showed at ICML that large language models identify text roles by style, not by tags. A forged chain-of-thought bypasses safety in OpenAI, Anthropic, Alibaba and DeepSeek models — and the problem may be unsolvable at the model level.

Aktualności31 lipca 2026

A self-replicating prompt-injection worm in Word's Copilot

Researcher Håkon Måløy detailed a prompt injection that, via Copilot in Microsoft Word, copies hidden instructions into new files — turning them into fresh carriers. Disclosed to Microsoft 144 days ago, still without a full fix.

Aktualności21 lipca 2026

GPT-Red: OpenAI trained an AI to break other AI models

OpenAI trained GPT-Red — an AI model specialized in breaking other models via prompt injection. GPT-Red achieved 84% attack success versus 13% for human red-teamers and was used to harden GPT-5.6 Sol to a 0.05% attack success rate.

Aktualności14 lipca 2026

Context bombing: using AI guardrails against hacking agents

Tracebit developed "context bombing" — placing prompt injections alongside AWS secrets to shut down attacking AI agents. Across 152 test sessions, admin privilege escalation dropped from 57% to 5%.

Aktualności10 lipca 2026

HalluSquatting: turning LLM hallucinations into botnets

Researchers detailed HalluSquatting — an attack that exploits LLMs’ tendency to hallucinate repository identifiers. By registering predictable fabricated names, attackers can mass-infect coding agents like Cursor and GitHub Copilot.

Aktualności4 lipca 2026

New attack on AI browsers: telling the LLM 2+2=5 bypasses safety guardrails

Researchers demonstrated that feeding an LLM-based browser a false fact is enough to make it abandon its safety guardrails. The BioShocking attack worked on six AI browsers.