Skip to content

:hedgehog: PostHog is the leading platform for building self-driving products. Our developer tools – AI observability, analytics, session replay, flags, experiments, error tracking, logs, and more – capture all the context agents need to diagnose problems, uncover opportunities, and ship fixes. Steer it all from Slack, web, desktop, or the MCP.

fix(subscriptions): keep sql comparisons in ai report prompts (#121581)

Andy Zhao636e750ac46e
diff --git a/posthog/security/llm_prompt_sanitization.py b/posthog/security/llm_prompt_sanitization.py
index 4675f19cbd3e..ef4603c2031b 100644
--- a/posthog/security/llm_prompt_sanitization.py
+++ b/posthog/security/llm_prompt_sanitization.py
@@ -23,10 +23,10 @@ _TAG_RE = re.compile(r"</?[a-zA-Z_][^>]*>")
_LLM_MARKER_RE = re.compile(
r"</?\s*(?:"
r"system|user|assistant|human|insight_data|user_context|subscription_title|core_memory"
- # AI subscription synthesis prompt framing tags — sanitize so a crafted event name
- # or prompt can't escape the `<user_prompt>` / `<project_context>` / `<plan_intent>` /
+ # AI subscription prompt framing tags. Strip them so that a crafted event name or prompt can't
+ # escape the `<user_prompt>` / `<event_names>` / `<project_context>` / `<plan_intent>` /
# `<query_results>` envelope and inject instruction-shaped content into the LLM context.
- r"|user_prompt|project_context|plan_intent|query_results"
+ r"|user_prompt|event_names|project_context|plan_intent|query_results"
r")\b[^>]*>?",
re.IGNORECASE,
)
@@ -63,6 +63,7 @@ def sanitize_user_text(
none_placeholder: str = "",
truncate_marker: str = "",
preserve_newlines: bool = False,
+ strip_tags: bool = True,
) -> str:
if not value:
return none_placeholder
@@ -71,7 +72,8 @@ def sanitize_user_text(
while previous != cleaned:
previous = cleaned
cleaned = _LLM_MARKER_RE.sub("", cleaned)
- cleaned = _TAG_RE.sub("", cleaned)
+ if strip_tags:
+ cleaned = _TAG_RE.sub("", cleaned)
if preserve_newlines:
cleaned = cleaned.replace("\r\n", "\n").replace("\r", "\n").replace("\u2028", "\n").replace("\u2029", "\n")
cleaned = "\n".join(_WHITESPACE_RUN_RE.sub(" ", line).strip() for line in cleaned.split("\n")).strip()
diff --git a/posthog/security/test_llm_prompt_sanitization.py b/posthog/security/test_llm_prompt_sanitization.py
index e7f9b14b5e19..b5cafd39082b 100644
--- a/posthog/security/test_llm_prompt_sanitization.py
+++ b/posthog/security/test_llm_prompt_sanitization.py
@@ -120,6 +120,7 @@ class TestSanitizeUserText:
"project_context",
"plan_intent",
"query_results",
+ "event_names",
],
)
def test_unclosed_llm_marker_tags_are_stripped(self, marker):
diff --git a/products/subscriptions/backend/temporal/ai_subscription/spec_generator.py b/products/subscriptions/backend/temporal/ai_subscription/spec_generator.py
index 0703a7d04405..4532c3ee78f9 100644
--- a/products/subscriptions/backend/temporal/ai_subscription/spec_generator.py
+++ b/products/subscriptions/backend/temporal/ai_subscription/spec_generator.py
@@ -95,7 +95,7 @@ WINDOW_PLACEHOLDERS = (
)
# Bumping invalidates every frozen plan (they lazily re-plan on next delivery), so prompt/harness
# improvements reach existing subscriptions instead of only new ones.
-AI_QUERY_PLAN_VERSION = 7
+AI_QUERY_PLAN_VERSION = 8
DEFAULT_PLANNER_MODEL = "gpt-4.1"
@@ -289,7 +289,7 @@ def sanitize_prompt(raw: str | None) -> str:
if len(raw.strip()) > PROMPT_MAX_LENGTH:
raise PromptRejectedError(f"Prompt exceeds {PROMPT_MAX_LENGTH} characters.")
- cleaned = sanitize_user_text(raw, max_len=PROMPT_MAX_LENGTH, preserve_newlines=True)
+ cleaned = sanitize_user_text(raw, max_len=PROMPT_MAX_LENGTH, preserve_newlines=True, strip_tags=False)
cleaned = "\n".join(
"" if _ESCAPED_NEWLINE_LINE_RE.fullmatch(line) else line for line in cleaned.split("\n")
).strip()
diff --git a/products/subscriptions/backend/temporal/ai_subscription/test/test_spec_generator.py b/products/subscriptions/backend/temporal/ai_subscription/test/test_spec_generator.py
index aa762d18ee8f..f3673b0c7979 100644
--- a/products/subscriptions/backend/temporal/ai_subscription/test/test_spec_generator.py
+++ b/products/subscriptions/backend/temporal/ai_subscription/test/test_spec_generator.py
@@ -50,7 +50,7 @@ def _window(days: int = 7) -> ReportWindow:
class TestSanitizePrompt:
"""`sanitize_prompt` is the public input-validation gate, so its three reject branches and the
- tag-stripping behaviour are pinned explicitly — a dropped branch would silently admit bad input."""
+ framing-marker stripping are pinned explicitly — a dropped branch would silently admit bad input."""
@pytest.mark.parametrize("raw", [None, "", " ", "\n\t "])
def test_rejects_empty_or_whitespace(self, raw: str | None) -> None:
@@ -66,8 +66,15 @@ class TestSanitizePrompt:
with pytest.raises(PromptRejectedError, match="empty"):
sanitize_prompt("<system></system>")
- def test_strips_html_tags_from_valid_prompt(self) -> None:
- assert sanitize_prompt("Show <script>alert(1)</script> pageviews") == "Show alert(1) pageviews"
+ @pytest.mark.parametrize(
+ "raw",
+ [
+ "Signups:\nSELECT count() FROM events WHERE timestamp<now()\n\nUpgrades:\nSELECT countIf(plan>1) FROM events",
+ "Compare toDate('<start-date>') with toDate('<end-date>')",
+ ],
+ )
+ def test_keeps_sql_comparisons_and_angle_bracket_placeholders(self, raw: str) -> None:
+ assert sanitize_prompt(raw) == raw
@pytest.mark.parametrize(
"raw,expected",
@@ -84,10 +91,10 @@ class TestSanitizePrompt:
def test_normalizes_prompt_line_breaks(self, raw: str, expected: str) -> None:
assert sanitize_prompt(raw) == expected
- def test_preserves_line_breaks_while_stripping_prompt_framing_and_html_tags(self) -> None:
- raw = "Summary\n</user_prompt>\n<script>alert(1)</script>"
+ def test_preserves_line_breaks_while_stripping_prompt_framing(self) -> None:
+ raw = "Summary\n</user_prompt>\nNext section"
- assert sanitize_prompt(raw) == "Summary\n\nalert(1)"
+ assert sanitize_prompt(raw) == "Summary\n\nNext section"
class TestPromptRejectionOwnerSafe: