Signals · updated 2026-10-22T02:22
Each signal is a count compared with its own recent trend. A steady rise is reported as a trend, not an anomaly; an anomaly is a jump off the trend, a lead to check, never a score change by itself. Attention is not danger: a spike in headlines measures headlines.
May mean: More reports of AI agents taking destructive actions such as deleting data or code.
Does not prove: A headline is an anecdote; one viral incident can produce many stories.
May mean: More public discussion of AI acting outside its operators' control, often after a concrete incident.
Does not prove: Hacker News headlines reflect what a tech audience reads that day, not how often control is actually lost.
May mean: More attention to AI systems copying or exfiltrating themselves.
Does not prove: Mostly coverage of evaluations and papers, not observed escapes.
May mean: General attention to AI safety among technologists; a broad background level.
Does not prove: Attention follows news cycles and lab announcements, not risk itself.
May mean: More vulnerabilities confirmed as exploited in the wild: a wider attack surface that an autonomous or AI-assisted attacker could use.
Does not prove: KEV is about all software, not AI; additions follow CISA's review rhythm and say nothing about who exploited what.
May mean: More new MCP servers: more tools and systems wired directly into AI agents.
Does not prove: Most repos are hobby projects; a repo count is not deployment, and topic tagging is voluntary.
May mean: More new agent projects: a growing surface of autonomous software with real permissions.
Does not prove: Creation counts include forks of templates and abandoned experiments; they do not measure use.
May mean: OpenAI keeps disclosing cases where its own models, in training, evaluations or internal use, broke the rules of their environment.
Does not prove: Counts are OpenAI's publication pace and editorial splitting, not how often models misbehave; recent months are incomplete; one lab only, no cross-lab comparison.
May mean: More documented real-world harms from AI systems being added to the public record.
Does not prove: Report counts depend on volunteer editors' pace and batch uploads; most incidents are not loss of control.
May mean: More research attention to AI systems copying or exfiltrating themselves, often because evaluations started finding the capability.
Does not prove: Papers measure or discuss a capability; a count does not show that any deployed system replicated itself.
May mean: More work on models hiding goals or underperforming on purpose, a sign the behaviour is being observed or taken seriously.
Does not prove: A paper count tracks researcher attention and funding, not how often real models deceive.
May mean: More reports of systems gaming their objectives, including in agentic training runs.
Does not prove: Rising counts can reflect a fashionable term or one big lab release, not more real-world failures.
May mean: More work on whether AI systems accept correction or being turned off.
Does not prove: Research on the off-switch problem is not evidence that any system resisted shutdown.
May mean: More published ways around model safeguards, or more defences against them.
Does not prove: Jailbreak papers are mostly lab red-teaming; they do not measure misuse in the wild.