Tonight's September 13 catch-up begins with research that put the question of how humans and AI should work together to a two-year experiment. The early-evening slot covered a BBC report on anxieties on the ground in the AI industry; tonight brings measured data from the classroom, agreement among industry leaders on an oversight regime, and security news for developer tools. The MCP authentication bypass in older LiteLLM versions has been confirmed exploited in the wild, and in the Japanese-language sphere, a verification piece that re-measured the effect of a much-discussed Claude Code plugin across 42 cells also drew attention.
Students Banned From AI Ranked Last Two Years Running — A Law School's Two-Year Comparative Experiment
According to The Decoder, a law professor spent two years running an experiment comparing student grades under three conditions: a group banned from using AI, a group that used AI with no guidance, and a group given structured training. The result: the group that avoided AI landed at the bottom two years in a row.
What stands out is the researcher's own words. The professor wrote “I was wrong,” admitting to having believed until then that unguided AI does more harm than good. The results can be read as showing that AI itself doesn't impede learning — outcomes shift depending on how its use is designed — and that reading is likely to stir debate.
That said, we should stop short of firm conclusions: the data collected this time doesn't tell us which group performed best, or the details of the experiment's scale and grading methods. Even so, the fact that a simple “ban AI” rule failed to show results in the face of a two-year span and hard numbers is weighty material for anyone thinking about rule-making in education.
Altman, Musk, and Hassabis Back Amodei's Call for “Independent Oversight” — OpenAI's IPO Comes in 2027
The debate over slowing AI, which ran on through today, has gained a new development. According to The Decoder, Sam Altman (OpenAI), Elon Musk (xAI), and Demis Hassabis (Google DeepMind) have at least partially endorsed Anthropic CEO Dario Amodei's call to slow AI development and introduce independent oversight.
Daytime coverage framed it as “rivals also agree on slowing down,” but this time the three figures who lead the industry have put their names to it explicitly. Alongside that, Altman stated clearly that OpenAI's IPO will be pushed to 2027, citing safety concerns. This puts more concrete backing behind this morning's “IPO shelved” reports.
It has also been reported that, in a Fortune interview, Altman spoke about his fear of AI's extinction-level risk. The details of his remarks couldn't be confirmed from the data collected this time, but the simple fact that an industry leader keeps speaking frankly about risk captures the drift of these past few days.
How Will the EU Answer the “Extinction Warnings”? Politico Reports From the Inside
As warnings from US industry figures continue, the regulators' side is coming into view. In an article titled “Keep calm and regulate on,” Politico Europe reports from inside the EU on how the bloc is responding to warnings about AI extinction risk.
The title alone conveys the EU's posture: don't get swept into panic, keep steadily operating the existing regulatory framework. After this week's warning from Amodei and the industry's slowdown agreement, how will regulators actually move? We couldn't verify the article's detailed contents from the data collected this time, but as the regulators' “answer” to the industry's “voice,” this is a report worth watching.
MCP Auth Bypass in Older LiteLLM Versions — Passes With a “Nonsense Key,” and Exploitation Is Confirmed
Critical security information for developers has also landed. According to an article on Qiita, a vulnerability in older versions of LiteLLM — widely used as an LLM API gateway — allows authentication for MCP (Model Context Protocol) sessions to be bypassed. An attacker without a valid API key can establish an MCP session with an arbitrary Bearer token and call MCP tools connected to in-house AI systems.
What makes it serious: on September 2, 2026, US CISA published this vulnerability as confirmed to have been exploited in the wild. This is not a hypothetical risk — attacks are already running. In environments that permit file operations or external service integrations over MCP, checking for exposure is now urgent.
Updating to the latest version is the baseline response, but the specific patched version and workaround details couldn't be confirmed from the data collected this time. If you run LiteLLM, check the official advisory and release notes directly.
Diff Agent Skills Against Their GitHub Source Before Installing — Demonstrated With a Lark Skill
When you install a “skill” for a coding agent, do you check whether the files the distributor hands you match the ones in the public repository? A Qiita article opens with “I suspect most people don't. We didn't either,” and walks through actually diffing ByteDance's Lark (Feishu) document skill “lark-doc.”
A skill is nothing less than the instructions and tool definitions you hand an agent, so if what the distributor ships diverges from the public code, something will behave differently from what you expect. It's the AI-agent edition of the “verify what's inside the artifact” discipline that has become second nature for npm packages and Docker images.
What discrepancies actually turned up in lark-doc this time, and how they were handled, couldn't be confirmed within the collected data. But simply proposing the habit of “diff a skill before you install it” is a point that should land with the many developers wiring agents into their work.
Re-Testing the “54% Fewer Diff Lines” Claim Across 3 Models and 42 Cells — ponytail's Effect Is 9–25%
In the Japanese-language sphere, a Zenn verification article that measured the actual effect of “ponytail,” a plugin that teaches Claude Code to “write like a lazy senior developer,” is drawing attention. The plugin's README advertises “54% fewer diff lines and 20% lower cost,” but the author re-measured it using a small Express app of their own and 7 tickets, running 42 cells in total across the three models Haiku, Sonnet, and Opus on headless Claude Code.
The result: reductions of just 9–25%, and “most of what was cut was line breaks.” Cost dropped for Opus alone, so taking the README's numbers at face value looks difficult. For a popular plugin said to have gained 11,730 stars this week alone, the arrival of independent measured data is no small thing.
Results should vary with benchmark conditions (the app's scale, the tickets' content), so this verification doesn't mean “the plugin is worthless.” Even so, the stance of “measuring a claimed effect in your own environment to confirm it” feels like a baseline that applies to every AI tool adoption decision.