self.md radar — 2026-11-07
TikTok’s next similar video
the European Commission’s new TikTok finding names vulnerable adults as well as minors
it concerns repeated recommendations about sadness, isolation and body image
1. TikTok’s repeated recommendations draw a DSA finding
in its November 6 notice, the Commission describes a familiar sequence: interact with a type of video, receive more of it
endless exposure to similar material can harm mental health, it says, particularly for vulnerable adults and minors
the preliminary finding covers both TikTok’s assessment of that risk and its mitigation measures
this is a preliminary breach finding, not a final infringement decision or an announced fine
the short notice does not prescribe a replacement feed
reading: the Commission’s finding on TikTok’s For You feed
2. each Replit sandbox gets a Linux kernel
Replit’s old containers shared the host kernel
a reachable flaw there could expose other workloads, credentials or host network endpoints
the company says it has finished rolling out a replacement built on KVM and Cloud Hypervisor
containers still divide processes inside a sandbox; a microVM now separates that sandbox from the host
compromising the guest kernel alone does not reach the host
Docker can receive privileges inside the guest that Replit could not safely expose in the old runtime
credentials take a separate route
managed connectors use a broker that authorizes a request, removes caller-supplied authentication headers and attaches the upstream credential
the workload can use the service without receiving its account API key
the egress qualification is easy to miss: host-side filtering applies to sandboxes with an egress policy
the post does not say every sandbox has a destination allowlist
reading: Replit’s sandbox design and security mechanisms
3. Claude’s science sessions lean heavily toward data analysis
48.1% of the Claude Code research sessions in Anthropic’s new study concern data analysis
ideation and hypothesis generation account for 2.1%
in a separate survey of 528 scientists, respondents rank ideation first and data gathering second for producing important results
the authors warn that session classification may undercount ideas developed during other tasks
the report also models research output rising as much as 15-fold in some fields if AI can perform all non-physical research work
that is an illustrative scenario, not an observed acceleration
its assumptions include AI-agent time costing $30 an hour and taking 0.2 hours per researcher-hour saved
reading: Anthropic’s usage data, survey and acceleration assumptions
sample: 207,049 scientific sessions on consumer plans, August 2–September 30, 2026
survey respondents came from applicants to Anthropic’s AI-for-science and related grant programs in the US and Canada