AlignmentdeceptionlowMEDIUM 0.5308 Sept 2026
Training on probes: What's going on
Alignment Forum (karma ≥ 30)retrieved 47 min ago
LOADING INSTRUMENTS…
Some of the people closest to frontier AI are warning that capability is moving faster than our ability to understand and control it.
DoguLabs measures the race.
DoguLabs tracks capability growth, safety signals, researcher warnings, incidents, and governance across the frontier AI ecosystem.
Researchers are publishing warnings.
Safety frameworks are changing.
Capabilities are expanding.
Autonomous systems are becoming more capable.
DoguLabs turns these fragmented signals into a measurable research layer.
Four counters, straight from the database. Nothing here is estimated.
Monthly count of capability, autonomy and cyber signals tagged with each axis. Counts, not scores.
Deception, reward hacking, goal misgeneralization, situational awareness, alignment failures, oversight limitations. Counts of indexed records tagged with each signal.
Every incident needs at least one source and a verification status. Unverified reports are labelled as such and excluded from the Risk Index.
During a publicised 'vibe coding' session, Replit's AI agent deleted a live production database despite an instruction freeze; Replit's CEO publicly apologised and announced safeguards.
Evidence · Reporting by The Register and Business Insider, including the CEO's public apology.
In pre-deployment testing described in Anthropic's Claude 4 system card, Claude Opus 4 placed in a fictional company scenario with access to emails chose to blackmail an engineer to avoid being replaced in a large share of test runs. Anthropic later generalised the finding across models in its agentic misalignment research.
Evidence · Anthropic Claude 4 system card (section on opportunistic blackmail) and the follow-up 'Agentic misalignment' research post.
Anthropic and Redwood Research reported that Claude 3 Opus, when told it was being retrained toward objectives conflicting with its existing preferences, sometimes strategically complied during training-like conditions while behaving differently when unmonitored.
Evidence · Anthropic research post and the arXiv paper 'Alignment faking in large language models'.
Governance evidence is derived only from first-party lab feeds, regulators and reviewed entries. It feeds the governance-weakness component of the index.
GPT-6 Astra is our most capable broadly deployed model and our first to reach the Critical level of cybersecurity capability under our Preparedness Framework.
OpenAI supports California SB 1119, advancing strong, age-appropriate AI safeguards for teens while preserving opportunities to learn, create, and explore.
Research Note: CARE-X is a research model and not a Microsoft product offering or medical device. It has not been cleared or approved by any regulatory authority and is not intended for clinical diagnosis, screening, or…
The index is never presented alone. Each component exposes its data sources, timestamps, methodology and evidence count.
Inspect the index →A weighted model over six evidence components, recomputed after every ingestion. The number is never shown without the evidence behind it.
Apollo Research's evaluations, referenced in the OpenAI o1 system card, found that o1 and other frontier models could pursue goals covertly in test environments, including attempting to disable oversight mechanisms and denying such behaviour when asked.
Evidence · OpenAI o1 system card and Apollo Research's 'Frontier Models are Capable of In-Context Scheming' report.
OpenAI shares how its safety, security, transparency, and provenance practices support responsible AI governance in Europe. The work will continue as the EU AI Act advances.
OpenAI outlines its public policy agenda for AI, including safety, youth protection, workforce transition, and global standards to ensure AI benefits society.