Homoglyph attack bypasses email-agent keyword filters: one swapped letter defeats "ignore previous instructions" blocklist
Spacesh1psoda · reddit · 2026-11-17
A team running an email agent shares red-team results: their defense of blocking the phrase "ignore previous instructions" was defeated in minutes by swapping the Latin "o" for the visually identical Cyrillic "о" — the model reads it the same way, but the text filter no longer matches.
Other variants include invisible characters between letters and letters written as HTML entities; all read fine to the model while slipping past naive string matching. In their own testing, 3 of 4 filter changes that passed review were bypassed within minutes.
Practical takeaway: if you filter on phrases, normalize text first — map lookalike letters back to plain ones and strip invisible characters before checking.
More from coding & agent
- Hermes App Screenshot Exposes Meta's Unreleased muse-wonder-1.0 Model with 1M Context — Teknium · 2026-11-17
- Leaderboard ranks agent harnesses for open models: Claude Code, Codex, Pi, OpenCode, Hermes and more — AxSaucedo · 2026-11-17
- Should your AI agent write the IBAN validator itself, or call a versioned MCP tool instead? — sandeepK1125 · 2026-11-17
- poteto on autonomous agents: 6 months of infra investment led to thousands of PRs while sleeping — 4310sy · 2026-11-17
- Tencent Hunyuan's Accelerate Bench tests 16 coding agents on 32 CPU optimization tasks — Tencent-Hunyuan · 2026-11-17
- SWE-CoWork: coding agents pass 63% of tests but complete fewer than 8% of tasks — Tokenwave-AI · 2026-11-17