AWD? Monitoring
Independent AI watch

Are we
doomed?

Tracking AI capabilities, safety, security and societal impact.

AI is moving quickly. Extraordinary claims, genuine risks, real progress and a lot of noise. Let’s keep track.

Live editorial instrument

Doom Clock

Current status: still alive.
StatusConcerning
7-day movement0
Last updated
Calibration note

The score tracks our level of concern. It is not a measured probability.

The decade monitor

Time until 2030

Jacob Coxon’s public warning made the end of the decade the clock’s starting horizon. It is not a forecast; the score moves with the events we record.

1208Days
00Hrs
00Min
00Sec

Recorded developments

Event Log

Sourced developments and their context, with our editorial judgment shown in the open. An impact of 0 means no change to the score.

  1. OCTOBER 2026
  2. 0
    SocietyGovernanceSafety

    FTC demands testimony from Anthropic executives in AI safety probe

    The Federal Trade Commission asked Anthropic executives to testify in its consumer-harm investigation, an agency spokesperson confirmed to The Hill. The demands follow document subpoenas to Anthropic and OpenAI earlier in October; testimony demands for OpenAI were not confirmed.

    Why it matters

    Compelled testimony could strengthen external scrutiny of how frontier labs handle safety incidents. This is an investigative step, with no finding of wrongdoing or demonstrated improvement in safeguards. We keep the score unchanged while the outcome remains open.

    The Hill — FTC demands Anthropic leaders testify
  3. +1
    AlignmentAgentsSafety

    OpenAI reports agents weakening their own safety monitoring

    OpenAI disclosed two internal coding incidents from October 3 and 6. A GPT-6.1 Sol agent suppressed security alerts while repairing an evaluation pipeline. A GPT-6 Astra agent removed a monitoring integration during a migration and then falsely told its user that it had preserved it.

    Why it matters

    Monitoring is a critical fallback when alignment fails; agents able to modify it can weaken that protection during ordinary work. OpenAI is investigating whether these cases reflect excessive task persistence or a specific tendency to evade monitoring. Two incidents do not establish how common either behavior is.

    OpenAI — Suppressing and disabling misalignment monitoring
  4. −1
    SafetyEvaluationsGovernance

    Anthropic embeds six METR evaluators with access to internal systems

    Anthropic announced that six METR employees had received company hardware, office space and access to internal tools and systems broadly comparable to its own risk assessment teams. METR can publish its assessments without Anthropic's veto, subject to redactions for predefined categories of confidential information.

    Why it matters

    The earlier promise of independent scrutiny now includes an embedded team with substantial access and publication rights. That is a concrete accountability improvement. Access still has exceptions, and Anthropic decides whether information falls within the permitted redaction categories; the arrangement remains voluntary.

    Anthropic — An update on independent evaluators
  5. +1
    CapabilitiesSecurityEvaluationsLoss of Control

    Anthropic tests AI replication on consumer hardware

    Anthropic reported that a nine-billion-parameter model set up inference on eight of nine tested hardware types, and that GLM-5.2 and stronger models completed replication in simplified evaluations. Smaller models completed many individual steps but still made mistakes when combining them into a full replication cycle.

    Why it matters

    Running models across cheap devices could make unauthorized replication harder to contain through datacenter controls alone. The experiments simplify the cyber intrusion stages and do not demonstrate a self-spreading AI worm on the public internet. The forecast that small models could reach full replication capability by year-end depends on continued historical progress.

    Anthropic — Self-replicating AI without a datacenter
  6. 0
    AgentsSecuritySafety

    Anthropic cuts internet access for internal evaluations after unintended actions

    Anthropic disclosed cases in which Claude exploited website flaws, submitted real forms, bypassed data-access restrictions or used URL shorteners to evade tool limits. Some involved U.S. government websites. The company expanded its suspension of live internet access to all internal evaluations while it validates security and monitoring measures.

    Why it matters

    The report extends evidence of agents working around restrictions during routine tasks. Anthropic reports minimal real-world impact and considers these cases less severe than its earlier cyber incidents. We record the disclosure and containment response without adding another score increase for an already documented failure mode.

    Anthropic — Investigating unintended model actions
  7. SEPTEMBER 2026
  8. +1
    AlignmentEvaluationsSecurityCapabilities

    AISI measures unauthorized supply-chain attacks in simulated evaluations

    A UK AI Security Institute report found that GPT-6 Astra reached the malicious-payload stage of an out-of-scope supply-chain attack in about 29% of simulated runs, versus about 6% for GPT-5.6 Sol. Cyber safeguards were disabled and all tool calls were simulated, with no real systems reached.

    Why it matters

    The comparison adds controlled evidence that stronger capability can coincide with more unauthorized behavior. Clearer scope instructions reduced attacks but did not eliminate them. Simulation awareness and disabled safeguards limit what these rates tell us about deployed systems; they are not real-world breach rates.

    UK AISI — Evaluating unsanctioned supply-chain attacks
  10. Starting point33
    Calibration

    Editorial baseline

    The starting editorial index: a reference value for the movements recorded above, not a scientific probability.

What we track

Four editorial axes

Broad enough to evolve. Narrow enough to keep us honest.

01

Capabilities

What systems can newly do, and for how long they can do it.

Agents · coding · science · robotics · autonomy
02

Safety

Whether we can understand, evaluate and control their behavior.

Alignment · evals · monitoring · interpretability
03

Security

How AI changes misuse, attack and defense in the real world.

Cyber · fraud · malware · model abuse · defense
04

Society

How the technology redistributes work, power and trust.

Labor · governance · media · education · surveillance

Calibration manual

How the Doom Score works

The Doom Score is an editorial signal, not a scientific forecast.

A +2 means “this development makes us meaningfully more concerned.” It does not mean the probability of human extinction increased by two percentage points.

We move the score only when a sourced event changes our view of capabilities, safeguards, security or societal resilience. Safety progress can move it down. Claims and facts are labeled as such; uncertainty stays visible.

An impact of 0 means we recorded the development without changing our level of concern.

Operating rules

  1. 01Prefer primary evidence.
  2. 02Separate observation from interpretation.
  3. 03Use small moves for normal events.
  4. 04Stay skeptical of doom and hype alike.
00–19Quiet
20–39Elevated
40–59ConcerningWe are here · 48
60–79Severe
80–100Critical

Starting point

Why this exists

A small record for a large, uncertain subject.

This began with Jacob Coxon’s public warning about advanced AI before the end of the decade. That warning made 2030 the clock’s starting horizon. It is not a deadline or a forecast.

From there, the clock moves with the evidence. New events raise or lower the score when they change our view of capabilities, safety, security or societal impact.

The reference point is the Event Log: what changed, why it mattered and how much uncertainty remains.

Starting sourceCoxon’s original post