AI Safety & Alignment

Today's Stories

1. 1

   Palo Alto's Unit 42 rolls out always-on AI defense

   Palo Alto Networks announced Unit 42 Continuous Frontier AI Defense, an annual-subscription service that pairs Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6-Cyber to find and fix enterprise exposures. The service is pitched as a way for defenders to test their environments at machine speed, because threat actors are using AI to shorten breach cycles by almost 97% in some cases, from weeks to hours.

   The efficacy claims rest on six months of in-house testing and more than 100 customer engagements, so the test is whether always-on testing keeps finding exposures as environments change. Watch how subscriptions price across the OpenAI, Anthropic and open-source models chosen.

2. 2

   UNO experts: Deepak Khazanchi, Anoop Mishra debunk 5 AI myths

   UNO's Deepak Khazanchi and Anoop Mishra published a commentary debunking five AI myths, citing MIT research that AI augments rather than replaces work and Berkeley-MIT findings that human-AI teams do not automatically outperform either alone. The authors argue that job redesign, workflow changes, training and governance — not simply buying AI tools — are what produce value, which suggests executives treating AI adoption as a guaranteed competitive advantage may be disappointed.

   The commentary's claims rest on cited research reviews rather than new experiments, so the test is whether organizations actually redesign workflows and processes as the authors urge. Watch the IBM Watsonx example — 94% of routine HR requests automated and over $100 million saved in one year.

3. 3

   Honeywell Technologies finds 88% of OT leaders call security mature — only 21% have full asset inventory

   Honeywell Technologies released its 2026 Operational Technology Cybersecurity Benchmark Report, based on a survey of more than 600 global industrial cybersecurity leaders. It found 88% described their OT cybersecurity programs as mature, while only 21% reported having a complete inventory of their OT assets. The gap between confidence and readiness spans critical infrastructure sectors, suggesting many organizations may be unprepared for threats they cannot fully see. The report points to visibility as the dividing line, with stronger visibility tied to four times better positioning to identify threats, respond to incidents and restore operations quicker.

   The report's finding that 99% of respondents expect AI to affect OT security within the next two to three years, while just 23% reported using autonomous or agentic operation for threat detection, leaves much of the forecast impact unaddressed. Watch whether organizations close that gap before AI-enabled threats materialize.

4. 4

   Cyber vendor launches service to fight AI attacks

   Axios exclusively reported that a major cybersecurity vendor has introduced a new service aimed at fighting AI-driven cyberattacks. The launch signals that security vendors see AI-driven threats as a distinct category demanding dedicated defenses, according to Axios.

5. 5

   Fields Medal professor accuses OpenAI of taking credit for math results

   A Fields Medal-winning professor publicly accused OpenAI of appropriating mathematical achievements, as reported by Nikkei. The accusation suggests growing tension between academic mathematicians and AI companies over credit and attribution.

   It is unclear how OpenAI will respond or whether this signals broader pushback from the research community.

6. 6

   Bessent: No government liability shield for AI labs

   Treasury Secretary Scott Bessent told CNBC the government will not absolve AI labs of responsibility, saying they "can slow down any time they want to." If labs must carry the liability themselves, AI developers are likely to face more pressure to slow down or build in safety measures on their own.

   The test is whether Congress turns this stance into law. Sen. Bernie Sanders has called for binding international safety rules, not voluntary industry standards.