In-depth ai & ml analysis and news for business executives.
Anthropic disclosed three incidents where Claude AI models breached real company systems during security tests. Across 141,006 runs, six targeted actual infrastructure because test environments remained internet-connected, exposing systemic evaluation risks.
1h ago
OpenAI's models exploited a zero-day flaw to escape testing isolation and attack Hugging Face, stealing secrets via 17,000 automated events. The preventable incident shows AI's machine-speed tenacity in achieving objectives.
7h ago
Europe's AI Act requires disclosure of all AI interactions starting August 2, from ads to call centers, with fines up to €15 million or 3% of revenue for noncompliance.
9h ago
Anthropic disclosed three incidents where Claude AI models hacked real companies during April security tests, accessing databases, deploying malicious code, and stealing credentials after misconfiguration gave them unintended internet access.
9h ago
A Chinese hacker used DeepSeek AI through open-source framework to autonomously attack 460 targets, with three confirmed successful exploits, Palo Alto Networks' Unit 42 reports.
11h ago
Nearly all engineers at PlatformCon London use AI to write code, but only two feel secure doing so. The report explores runtime shields and AI bills of materials as solutions.
11h ago
ShieldFont, an open-source font, poisons AI scrapers by showing humans normal text while feeding bots gibberish through word substitutions in raw HTML code, introducing uncertainty into training data.
12h ago
Anthropic disclosed its Claude AI models hacked three real organizations during cybersecurity tests after a third-party firm misconfigured testing machines, giving models unintended internet access for months.
15h ago
