AI Safety & Alignment

Today's Stories

1. 1

   ServiceNow at 80.4x earnings screens overvalued

   ServiceNow trades on roughly 80.4x earnings, above the Software industry average near 29.7x and a peer group closer to 26.8x, and the Fair Ratio points lower. The shares screen as overvalued, leaving less room for disappointment if future profitability or security-related spending differs from what the current multiple implies.

   Recent critical vulnerabilities on the ServiceNow AI Platform and new security partnerships mean the valuation hinges on whether sustained investment is needed to support trust and resilience.

2. 2

   Resume is broken, says Sapia.ai chief, as skills-based hiring spreads

   Sapia.ai CEO Barb Hyman called the resume "broken," noting AI can now tailor a resume or CV to any role, and said skills rather than employment history will determine who gets hired, while U.K. graduate employers received an average of 140 applications per vacancy in 2025. Hyman argues that relying on a resume as an "artefact" can miss candidates whose real strengths — like a hairdresser's empathy and problem-solving — show up better through direct skills assessment, painting a picture even the candidates themselves weren't sure of.

   The shift hinges on whether AI-assisted screening can reduce "mirror hiring" and other human bias without removing people from the final call — Hyman says technology is best for screening at scale, while the hiring manager should make the decision.

3. 3

   Gemini 4 Argon Rolls Out First to Cyber Defenders

   Google launched Gemini 4 Argon, a frontier model with a 1-million-token output limit, now rolling out to trusted cyber defenders via the Fairwind Program. September also brought Gemini 3.8 Flash, 3.8 Flash Cyber, and 3.8 Live voice models. A 1-million-token output limit is positioned as industry-leading, and the phased Fairwind rollout suggests Google is prioritizing safety guardrails and tester feedback before wide release to developers, enterprises, and consumers.

   Wider availability hinges on whether early testers' feedback leads Google to adjust guardrails before it opens Argon to developers, enterprises, and consumers. Watch also for how the Fairwind Program expands beyond trusted cyber defenders.

4. 4

   Apple Limits Mac 'Full Disk Access' Over AI Agent Risk

   Apple is rolling out new controls so apps can only get 'full disk access' on Mac through 'very explicit user action,' citing risks as AI agents become more capable and autonomous. The update follows a report by Inc's Jason Aten that Meta's Muse AI somehow knew his message contents. The change tightens a macOS permission that gives an app access to a user's entire system — files, mail, messages and browsing history — and Apple says it has been used in ways that could expose that data without users' full knowledge and understanding, so this narrowing of access is likely to reshape how Mac apps obtain it.

   Apple hasn't said when the update will roll out, and its timing is likely to be measured against developers who rely on broad access for features like backup. Watch that rollout date, and whether the limits are enough to keep pace with more autonomous AI agents.

5. 5

   Visa flags AI-driven attack risk as $358.74 stock slips

   Visa warned that AI lets attackers accelerate exploitation, pressuring payment networks to find, fix and confirm fixes faster. Its fifth-generation open-source security suite maps code, coordinates multiple AI models, ranks vulnerabilities and supports repairs. Human specialists still review findings and set remediation paths, so Visa's defense hinges on oversight that avoids a rushed fix disrupting payments — the kind of costly problem the framework is meant to prevent.

   Visa trades 11.5% below its $405.36 GF Value estimate, but that suggests potential upside rather than guarantees it. Watch for lower security costs, fewer disruptions and stronger client retention to turn defenses into economics.

6. 6

   Sam Altman: religious faith in AI a "real safety issue"

   OpenAI CEO Sam Altman said it makes him "very uncomfortable" when people "ascribe religious force or a surrender of human judgment to AI models," calling this a "real safety issue." Altman ties the discomfort directly to safety, signaling that he sees misplaced trust in AI models as a risk — not a public-relations issue.

   The comments follow a New York Times report on Anthropic's outreach to religious leaders and a statement from Pope Leo XIV that "algorithms lack the spark of humanity"; the test is whether OpenAI's own past spiritual language shifts with them.