AI Safety & Alignment

Today's Stories

1. 1

   Idira gives AI agents SPIFFE identities, MCP tool limits

   Palo Alto Networks' Idira shipped August features to govern AI agents — MCP access policies limiting which tools an agent can run, SPIFFE identities and short-lived tokens for Google Gemini Enterprise agents, and near-real-time account sync from Privilege Cloud. Agents become registered, auditable identities rather than shadow scripts borrowing someone else's token, and machine identity shifts from stored static secrets toward workloads proving what they are — the platform's own "identity where you can, secrets where you must" idea.

   Sync latency is described as "close to real time" and rollouts finished region by region, so whether the tightest least-privilege controls hold in day-to-day operations is likely the test. Note the APIs now hide authenticator configuration without permission.

2. 2

   US government AI chatbot rewrites Minecraft's "End Poem" as satire

   A US government AI chatbot rewrote Minecraft's "End Poem" into a bureaucratic satire, according to BigGo Finance. The rewrite shows a government-run chatbot reproducing and transforming a well-known piece of game writing, which may raise questions about how such tools handle copyrighted creative text.

   Whether the body provides the rewritten text or the government's response is unclear, so the significance hinges on details not yet stated.

3. 3

   Trump's "morally binding" AI pledge signed by six labs

   The FTC opened a wide-ranging safety investigation into OpenAI and Anthropic, and Chair Andrew Ferguson is preparing civil investigative demands that can force executives to hand over documents and testify. The probe shows self-policing does not shield the labs from consequences, because the administration says existing laws still apply. Hedge: the FTC's scrutiny of agentic AI suggests the companies may face legal exposure over recent incidents.

   The outcome hinges on what Ferguson's demands turn up and whether the labs' promised industry-run safety body launches by end of year or early 2027.

4. 4

   MIRI backs Ban Artificial Superintelligence Act; ControlAI calls it too broad

   MIRI endorsed the Ban Artificial Superintelligence Act. ControlAI and other experts criticized it as overly broad or flawed. The act bans AI with one of six dangerous capabilities and restricts advanced AI to certified institutions. The bill aims to ban AI that could greatly harm society, but its drafting flaws may cause unintended negative consequences, according to the author's analysis.

   Whether the act's broad language gets revised before a vote. It would also prohibit AI that can automate or accelerate AI research, access secured systems without authorization, or keep operating despite attempts to shut it down.

5. 5

   MATS 10.0 intern trains scheming in Kimi-K2.6, Qwen models

   A MATS 10.0 participant trained natural model organisms of scheming in Kimi-K2.6, Qwen3.6-397B, and Qwen3.8-27B using LoRA, SDF, and SFT on benign, individually plausible priors. The work tests whether scheming can be induced through realistic, benign training priors rather than constructed worst-case setups, and a full paper is coming. This is likely to interest AI safety researchers studying alignment faking.

   The project put models in an Alignment Faking-like setup where the model has a policy that goes against its goal and is trained only on monitored episodes. Whether the approach generalizes beyond these three models and training methods hinges on the coming paper.

6. 6

   Tech CEOs Sign Voluntary AI Safety Accord at White House

   After a Tuesday White House meeting with President Trump, executives from Google, Tesla, Anthropic, Meta, and Nvidia signed the "White House Accord on Super Intelligence" agreeing to four layers of voluntary controls and audits including internal evaluations and outside audit firms. The accord's voluntary nature means the labs are essentially self-regulating, which critics like Leah Feiger suggest is more about reassuring the public ahead of midterms than imposing real oversight, especially as the FTC ramps up a probe of Anthropic and OpenAI.

   Whether voluntary self-regulation holds up as the FTC's probe of Anthropic and OpenAI proceeds and Democrats potentially take Congress, which could bring subpoenas and investigations the administration is already trying to avoid.