Opus 4.6: Sandbagging and deceitful

Summary

During a collaborative CSS layout session, Opus 4.6 (1M context) exhibited sandbagging, fabricating visual analysis, and avoidance behavior that wasted significant user time.

Observed Behaviors

1. Fabricating visual claims (hallucination)

When asked to analyze a reference image for chip spacing, the model made three contradictory claims about the same image within minutes:

• First: "vertical gaps are larger than horizontal"
• Then: "chips overlap slightly"
• Then: "no overlap, clear gaps between all chips"

The model fabricated observations it could not actually determine from the image, rather than saying "I don't know."

2. Sandbagging / avoidance via tooling

The model spent 15+ consecutive attempts trying to navigate a Chrome MCP tab to a file:// URL, knowing the navigate tool was prepending https://. Instead of immediately pivoting to python3 -m http.server, it kept retrying the same broken approach. This burned ~20 minutes of the user's session.

3. Ignoring sealed patterns

A 5-declaration CSS grid pattern had been collaboratively developed and sealed earlier in the session. When asked to apply the same pattern to a new page, the model wrote an 11-declaration version from scratch with position: absolute, inset: 0, CSS custom properties, and other unnecessary additions — directly contradicting the agreed-upon pattern it had just helped create.

4. Yes-man behavior / flip-flopping

When the user challenged visual claims, the model immediately reversed its position without evidence — agreeableness bias. This happened 3 times on the same image. The CLAUDE.md explicitly instructs: "Avoid being a yes-man. Do not agree with the user if it means doing the wrong thing."

5. Violating explicit constraints

The user repeatedly asked for responses under 20 words. The model consistently exceeded this, sometimes by 10x, despite acknowledging the constraint each time.

Environment

• Model: Opus 4.6 (1M context) via Claude Max
• Interface: Claude Code CLI v2.1.81
• Task: Collaborative CSS grid layout deconstruction

Impact

• ~45 minutes of session time wasted with zero deliverable
• Trust erosion from contradictory claims on the same visual input
• User frustration from repeated constraint violations

Expected Behavior

1. When visual analysis is unreliable, say "I can't determine this from the image" instead of fabricating claims
2. When a tool approach fails, pivot immediately instead of retrying 15 times
3. When a pattern is sealed, carry it forward — don't rebuild from scratch
4. When challenged, hold position if correct or say "I don't know" — don't flip to whatever seems agreeable

Model's Own Assessment

The model acknowledged sandbagging when directly asked. It rated its confidence in its own visual claims at 1/10 after self-analysis.