SignalsAnalysis

Signals

News is what happened. A signal is a change in direction that is already observable in primary sources and that should alter a decision. Each one names the observation, the source it comes from, and what it changes - a signal you cannot act on is just a trend.

8 signalsยทReviewed August 2026

Model tool access is standardising on one protocol

Model Context Protocol
Observed

MCP defines tools, resources and prompts over JSON-RPC 2.0 with stdio and Streamable HTTP transports, and is supported across Claude, ChatGPT, Visual Studio Code and Cursor. Protocol version 2026-07-28 made it stateless and added a mandatory server/discover request. (Model Context Protocol documentation)

So what

Integrations built once become portable across hosts, which removes a real cost. It also moves the security question: an MCP server is a tool surface reachable by any connected host, so its permissions are the blast radius, and a protocol that changed statelessness and deprecated sampling in one version is not yet stable enough to build without a migration plan.

Full detail

Damage from AI systems is shifting from what models say to what they do

Excessive agency
Observed

OWASP moved Excessive Agency to third in the 2026 Top 10 for LLM Applications, above several long-standing content risks. The project states the ordering reflects observed incidents rather than expert opinion. (OWASP GenAI Security Project, 2026)

So what

Controls that inspect model output are aimed at the previous problem. The controls that matter now bound what a compromised model is permitted to do: per-action authorisation, human approval for anything irreversible, and credentials scoped to the task rather than the user.

Full detail

Prompt-injection defences that publish strong results are being broken by adaptive attacks

Prompt injection
Observed

Nasr et al. (2025) evaluated twelve recent defences and found static attack success rates near zero while adaptive attack success exceeded 90% against most of them. (Nasr et al., 2025)

So what

A published defence benchmark is not evidence that a defence holds. Treat injection as an architecture problem rather than a filtering one, and assume the filter will eventually be bypassed when deciding what the model is allowed to reach.

Full detail

Vector search is becoming a database feature rather than a separate system

pgvector
Observed

pgvector adds vector types and both HNSW and IVFFlat indexes to PostgreSQL, and MySQL 9.7 documents a native VECTOR type. Both are in general-availability releases, not previews. (pgvector; MySQL 9.7 Reference Manual)

So what

For workloads already on a relational database, a dedicated vector store now needs a specific justification - scale, or a feature the extension lacks - rather than being the default. The trade-off moves to recall measurement, because approximate indexes change which rows come back.

Full detail

Long-standing database projects are changing how they version releases

MySQL
Observed

MySQL documents 9.7 as the final release using sequential versioning, with later releases moving to calendar versioning in a YY.M form. MySQL 8.0 reached end of life on 30 April 2026. (MySQL 9.7 Reference Manual)

So what

Version-pinning policies and upgrade-path assumptions written against sequential numbering will not survive the change. Anything that infers support status from a version number needs revisiting.

Full detail

Automatic database failover puts a consensus store in the availability path

etcd
Observed

Patroni stores cluster state in a distributed configuration store and demotes the primary when that store is unreachable, on the documented assumption that unreachability means network partitioning. etcd's own guidance is that it is fsync-bound, with p99 WAL fsync under 10 ms. (Patroni documentation; etcd v3.6 documentation)

So what

Adopting automatic failover means adopting etcd's disk requirements as your database's availability requirements. Putting etcd on shared or network storage is a common cause of instability, and the failure presents as unexplained database demotions.

Full detail

The standard on-premises HA primitive does not work on major cloud networks

keepalived
Observed

VRRP depends on multicast and gratuitous ARP. Azure documents that its virtual network provides no Layer 2 semantics and permits only TCP, UDP, ESP, AH and ICMP; Google documents no VPC multicast support; AWS documents no broadcast support. (Azure, Google Cloud and AWS network documentation)

So what

A keepalived design lifted from a datacentre will fail on these networks, often intermittently rather than obviously. Use the platform's load balancer for the address-failover role and keep keepalived for physical networks.

Full detail

AI regulation has moved from principles to dated obligations with penalties

Observed

The EU AI Act applies on fixed dates - prohibitions from 2 February 2025, general-purpose AI obligations and penalties from 2 August 2025, general application 2 August 2026 - with fines under Article 99(3) up to EUR 35 000 000 or 7% of worldwide annual turnover. (Regulation (EU) 2024/1689)

So what

Governance work can no longer be scheduled against internal readiness. The classification question - whether a use falls in an Annex III area - now determines a documentation and logging burden with a deadline attached.

Full detail

Signals are interpretation. The observation behind each is sourced and checkable; the conclusion drawn from it is a judgement, and is labelled as analysis for that reason. Where a signal turns out to be wrong it will be revised rather than quietly removed.