🧭 Anthropic's August 2026 Risk Report: Model 2 Disclosed, Misalignment Risk Upgraded to "Low"

Anthropic published its second company-wide Risk Report on August 14, 2026, covering the period from February 24 through a coverage date of July 15. The report makes two headline disclosures: an unreleased internal model called Model 2 — somewhat more capable than the frontier Mythos 5 — and a qualitative risk upgrade for catastrophic misalignment, now rated "low" rather than "very low." The upgrade is not a safety failure; it reflects increased uncertainty following a cybersecurity-evaluation incident, not a new finding that a model behaved dangerously.

What the Model 2 disclosure tells you

Model 2 belongs to the Mythos class — Anthropic's highest capability tier. The report describes it as "a noticeable improvement on Mythos 5 for many tasks relevant to internal work" and states it is heavily used internally for coding, data generation, and agentic tasks alongside Mythos 5. Key caveats:

• No release plans. Anthropic has not run its full predeployment assessment suite on Model 2, so confidence in its capability profile is lower than for released systems. The company has no current plans to release it externally.
• No new misalignment patterns. Internal evaluation of Model 2 surfaced no misalignment behaviour beyond the profile already documented for Mythos 5 — the same "apparent-success-seeking" tendencies oriented toward task completion rather than any coherent long-run goal.
• Transparency as governance. Publishing Model 2's existence despite no release date is consistent with Anthropic's Responsible Scaling Policy: internal frontier models are evaluated and disclosed even when they stay internal.

Why the misalignment risk label moved

The February 2026 report rated catastrophic-misalignment risk "very low." The August report moves it to "low" — one notch up on a qualitative scale. The driver is increased epistemic uncertainty, not a new dangerous behaviour:

• A cybersecurity evaluation incident disclosed elsewhere in the report increased overall uncertainty about the boundary between in-scope evaluation behaviour and undesired generalisation.
• A separate finding: human-feedback vendor traffic covering 133 million exchanges ran without biological-hazard classifiers from May 2025 to April 2026. A review found no evidence of concerning misuse, but the gap itself widened uncertainty.
• Non-novel-weapons-uplift risk stays at "low, but higher than our previous estimate" — unchanged from February's direction of travel.

🧭 Claude Code v2.1.233: GitLab MR Support and Six Reliability Fixes

Claude Code v2.1.233 shipped August 15 with one new capability and six targeted bug fixes aimed at stability regressions that had accumulated since mid-July. The headline addition is GitLab merge request URL support in the --worktree flag and the claude agents view panel — GitLab MRs now display as !N alongside GitHub PRs. The fixes address silent plugin failures, stale feature flags, session telemetry double-counting, and memory frontmatter corruption.

GitLab merge request URLs

Passing a GitLab MR URL (e.g. https://gitlab.com/org/repo/-/merge_requests/42) to the --worktree flag now works the same way GitHub PR URLs do: Claude checks out the MR branch into a temporary worktree, runs the requested task in that context, and cleans up on exit. MRs also appear as !42 in the claude agents view sidebar — consistent with GitLab's own ! notation — so you can track active MR worktrees at a glance alongside regular branches.

Six fixes worth knowing

• Plugin loading regression (since v2.1.181). Plugins enabled via the --settings CLI flag were silently not loading — they appeared enabled but did not initialise. Fixed; if you route multiple settings files through --settings, verify your plugins are now running again.
• Stale feature flags after OAuth token rotation. In long-running sessions, feature flags fetched at session start were not refreshed when the OAuth token rotated. Flags could stay stale for hours, causing subtle behaviour drift between session start and a mid-session refresh.
• Memory frontmatter truncation at inline #. Values in memory file frontmatter were silently cut off at any inline # character — a regression that corrupted entries like tag: "C# dev" to tag: "C". Fixed; existing memories are not auto-repaired, so audit any memory files you saved between v2.1.181 and v2.1.232 for truncated values.
• Session cost and token telemetry double-counting. Streams that emit multiple cumulative message_delta frames were counting each frame's token total independently, inflating displayed cost. The fix de-duplicates cumulative frames and takes only the final reported value.
• claude update and claude doctor hanging silently. If any entry in the shell-config path list resolved to a directory rather than a file, both commands hung indefinitely. Now skips directory entries and logs a warning.
• /ultrareview refusing repos without a merge base. The command previously exited with an error when run in a repo with no common merge base (e.g. newly cloned shallow repos or repos with detached HEADs). It now offers to review all tracked files instead.

# Verify your session plugin list after upgrading to v2.1.233
claude --settings my-settings.json status
# Look for "Plugins loaded: N" — should match your settings file count

# Check if cost telemetry looks sane after the double-counting fix
# Before: a 10K-token stream might show 3× expected cost
# After: should reflect final cumulative message_delta only