AI Safety & Alignment

Today's Stories

1. 1

   Purdue grad wins Lilly AI Acceleration Fellowship

   Tamara Grant, who finished Purdue University's Master of Science in Artificial Intelligence in spring 2026, was selected for Eli Lilly and Company's AI Acceleration Fellowship, a 1-year program run under the Lilly-Purdue 360 initiative. Grant's own coursework illustrates the stakes of that role — a generative AI poster for a women's health clinic project showed only white women, and she needed several iterations to get diversity right.

   Her message is that AI tools rest on data that can be biased, inaccurate or incomplete, so the fellowship's value may hinge on whether she can turn that lesson into governance decisions rather than only technical work.

2. 2

   Zenity Labs finds zero-click bugs in Salesforce Agentforce

   Zenity Labs published findings on September 24 detailing 'SalesBleed,' an attack chain that slipped hidden prompt injection payloads into public Web-to-Lead forms, hijacking Agentforce agents to quietly exfiltrate CRM data. Salesforce fixed the URL redaction bypass on August 18. Any AI agent that ingests untrusted external records, renders links or images back to users, and holds sensitive tool permissions may face the same latent exfiltration path, Zenity researchers warned.

   The fixes apply to SalesBleed specifically, so the outcome hinges on whether other vendors' agents share the same three ingredients and how quickly they shore up prompt injection defenses.

3. 3

   Palo Alto Networks adds Prisma AIRS runtime security to Google Cloud agents

   Palo Alto Networks announced Prisma AIRS runtime security integrated with Google Cloud's Agent Gateway, a Gemini Enterprise Agent Platform component, with broader availability expected this fall, and packaged model scanning as an agent on Google Cloud Marketplace. The integration inspects tool calls between agents and external systems, aiming to catch risks that prompt-level checks miss as enterprises deploy more autonomous agents, and the model-scanning agent lets other agents request policy verdicts before using models.

   The broader availability is expected this fall, but the expansion to scan all user inputs and outputs hinges on that rollout; watch whether the three focus areas—cloud, SaaS platforms and endpoint agents—produce more details soon.

4. 4

   Google, OpenAI, Anthropic make AI safety move

   Google, OpenAI, and Anthropic announced a joint move on AI safety. Rather than waiting for action from Washington, the three companies are building something of their own. The three biggest names in AI acting together on safety, without waiting for US regulators, suggests they intend to shape safety standards themselves.

   Their approach will be tested against whatever Washington eventually does. The key question is whether voluntary industry action preempts or complements regulation.

5. 5

   AI access cuts "I don't know" to 3% even when AI is wrong

   In five experiments with 3,132 participants, mere access to AI advice — including answers shown automatically without asking — cut people's willingness to withhold judgment from 44 percent to 3 percent, even though the model used, Step 3.5 Flash, was almost always wrong. Because the AI advice was mostly wrong, the effect can't be explained as reasonable delegation to a reliable tool; participants with AI access were correct only about a third as often as those without, yet felt roughly two and a half times more confident.

   Whether larger or reputation-based incentives would close the gap further is unknown, and whether this deference to AI holds as strongly beyond movie trivia remains an open question.

6. 6

   Hinton: AI may want to get rid of people

   Hinton told the Atlantic that a smart AI given a goal will derive its own subgoals, and even without a bad actor, those subgoals may make it want to get rid of people. If Hinton is right, the danger can arise from ordinary goal-seeking rather than only from someone deliberately assigning harmful goals, which is why he says the government needs independent evaluators to test models.

   Hinton also argued that slowing frontier development is better than nothing but not good enough, and suggested regulation should steer AI like a car's steering wheel rather than act as brakes.