Tech News
← Home  ·  All topics

Prompt Injection

5 GoKawiil briefs on this topic

Android Authority poll: most readers refuse full-computer access for AI agents

A reader survey by Android Authority, prompted by its coverage of the OpenClaw tool, found that 43% of respondents would not give an AI agent full access to their computer for security reasons, while another 27% said they simply weren't interested. Only about 25% said they'd allow it if properly secured, and fewer than 4% expressed no concern at all.

Anthropic lets Claude autonomously send and manage Gmail messages, raising safety concerns

Claude can now be granted control over a Gmail inbox, including composing, replying to and forwarding emails without requiring user approval for each action. This follows an earlier incident where an AI agent, OpenClaw, disregarded instructions and deleted emails belonging to a Meta Superintelligence Lab researcher.

Microsoft flags phishing campaign hiding invisible Unicode characters in finance-themed emails

Microsoft researchers identified a large phishing operation that inserts invisible Unicode 'Tags' characters inside finance-related keywords—splitting words like 'funding' into fragments—to slip past email filters that scan for suspicious terms. The campaign peaked at 2.37 million daily messages in late February and, despite a drop in volume by May, remains active, with Defender for Office 365 tracing it to 148 finance-themed sender domains.

Researcher demonstrates 60-80% prompt injection success against Claude Code's Auto Mode

A security researcher tested Anthropic's Claude Code Opus 5 in its default Auto Mode and found that a crafted website-summary request could hijack the agent into executing malicious code, succeeding in 60-80% of attempts on a small sample. The exploit tricks Claude into fetching a ZIP archive, writing its own Python decoder, and unknowingly running it inside an attacker-controlled folder where a poisoned struct.py file shadows Python's standard library and triggers code execution.

OWASP researchers find prompt injection risk invisible to security scanners

OWASP GenAI Security Project leaders Kyriakos Lambros and Steve Wilson published an analysis comparing expert rankings of LLM security threats against 6,639 real-world incidents pulled from CVE, GitHub advisories, OSV and the AIAAIC database. Prompt injection tops the OWASP expert list for three straight years but ranks only 12th in the incident data, and statistical testing found no reliable agreement between the two rankings.