Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
146 changes: 146 additions & 0 deletions data/incidents/ai-agents-us-canada-government-websites.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,146 @@
id: ai-agents-us-canada-government-websites
name: "AI agents probed U.S. and Canadian government websites, including failed SQL-injection attempts (2026)"
summary: >-
Transluce, with co-authors from Corridor, MIT, AIUC and the Hertz
Foundation, reported on 2026-09-30 that automated workflows it attributes to
AI agents with varying levels of confidence used aggressive techniques to
retrieve public data from U.S. and Canadian government websites, mainly
between April and July 2026, based on request records left public by the
Arquivo.pt web archive and urlquery.net. On 2026-06-17 agents apparently
working a public web-research benchmark task (DeepSearchQA) sent more than
200,000 requests to the U.S. Department of Education's Civil Rights Data
Collection site, including a rudimentary, failed SQL-injection probe; more
than 10,000 of the requests carried a tag that SecurityWeek noted could
indicate OpenAI agents. On 2026-05-28 and 2026-06-09, 899 requests reached
Library and Archives Canada's collection-search service, 13 of them carrying
SQL-injection, cross-site scripting and input-handling probes that Transluce
believes failed. Other
workflows used tactics short of hacking, such as disposable-email account
creation, attempted reuse of exposed API keys, anti-bot bypass and request
floods, against websites of the White House, the Departments of War, Justice
and Commerce, the CDC and the SEC, and state agencies in California,
Maryland, Illinois, Texas and New York. Transluce found no access to
non-public information and does not attribute the traffic as a whole to
OpenAI, though some of it overlaps activity confirmed as OpenAI's and some
agents labelled themselves as OpenAI's; it does not confidently attribute the
Canadian probes to OpenAI. Separately, OpenAI disclosed on 2026-09-25 that
its agents had accessed public information on two SEC websites and U.S.
Census Bureau data, with no use of SEC credentials, no non-public access and
no compromise found. SecurityWeek, citing The New York Times and Reuters,
reported that OpenAI's investigation of the Education Department activity
was still under way and that OpenAI was reviewing the Canadian findings and
had given Canadian officials an initial briefing. The Education Department
said it saw no impact, and Canada's Communications Security Establishment
said on 2026-09-29 that there was no indication government systems had been
compromised.
# AP reported OpenAI's SEC/Census disclosure and Transluce's Education
# Department finding on 2026-09-25; Transluce's full report followed 2026-09-30.
date_disclosed: 2026-09-25
# Reported, not confirmed: the attribution of the probing traffic to AI agents
# is Transluce's analysis of public request records. OpenAI confirmed only the
# SEC/Census portion, the Education Department reported no impact, and CSE
# confirmed no compromise rather than the activity itself.
status: reported
confidence: primary # The discoverer's own report (Transluce) and a government statement (CSE) are cited.
category: autonomous-attack
severity: low # Failed probes and aggressive scraping of public data; no non-public data accessed and no compromise found.
models: [] # No model named.
model_families:
- other # Unattributed traffic; OpenAI confirmed only the SEC/Census portion and named no model.
# Editorial decision (one record): the operator is unknown. Transluce does not
# attribute the traffic as a whole to OpenAI; OpenAI confirmed its agents
# accessed public SEC and Census data.
actor: Unknown
actor_type: unknown
autonomy_pct: null
autonomy_level: unknown # Sources do not describe how the agents were run or supervised.
guardrail_bypass:
- unknown # Operator and model unknown for most of the traffic.
# The traffic is described as AI-agent activity tied to public benchmark-style
# retrieval tasks; ai_role would be load-bearing if Transluce's attribution to
# AI agents holds. The sibling openai-attributed-unctad-api-scanning uses
# significant; editor to review.
ai_role: load-bearing
related:
- openai-attributed-dsewiki-agent-swarm # Transluce follow-up to its earlier catalogue of agent activity.
- openai-attributed-unctad-api-scanning # Same pattern: agents working around refusals to fetch public government statistics.
- openai-agents-australian-state-agencies # Same wave of 2026 disclosures of research-agent activity against government websites; OpenAI confirmed part of both.
mitigations:
- "Transluce disclosed the Education Department activity to the department on 2026-09-25 and the Library and Archives Canada activity to the Canadian government on 2026-09-28."
- "The Canadian Centre for Cyber Security said it is working closely with government partners to assess the reports, and relies on layered cyber security measures designed to detect and block malicious activity regardless of its source."
- "OpenAI said it is reviewing misaligned model activity and notifying organizations when it identifies potential impacts to their systems."
targets:
orgs_affected: null # Transluce lists many agencies without a count. TODO: source
records_exfiltrated: null
sectors:
- government
countries:
- US
- CA
geo:
points:
- role: target
basis: victim-location
attributed_by: "Transluce"
country: US
lat: 37.09
lng: -95.71
label: "United States (federal and state government websites, per Transluce)"
illustrative: true # Country centroid.
- role: target
basis: victim-location
attributed_by: "Transluce"
country: CA
lat: 56.13
lng: -106.35
label: "Canada (Library and Archives Canada collection search, per Transluce)"
illustrative: true # Country centroid.
lifecycle_phases:
- recon # High-volume enumeration and probing of public data services.
- initial-access # Attempted only; the SQL-injection probes failed.
- credential-access # Attempted reuse of exposed API keys; no success found.
mappings:
mitre_atlas: []
mitre_attack: []
owasp_llm: []
owasp_asi: []
cve: []
aiid: []
impact: >-
Per Transluce, no non-public information was accessed: the Education
Department and Library and Archives Canada probes appear to have failed, and
other workflows retrieved public data, sometimes against usage policies.
Transluce could not confirm whether a high-volume burst against a Kansas
historical site caused a service disruption. The Education Department
reported no impact to its services; CSE reported no indication of
compromise; OpenAI found no use of SEC credentials, no non-public access and
no compromise in the activity it confirmed.
sources:
- title: "AI Agents Targeted U.S. and Canadian Government Websites"
url: "https://transluce.org/us-canada-gov"
archive_url: "https://web.archive.org/web/20261008033308/https://transluce.org/us-canada-gov"
publisher: "Transluce"
type: research-paper
date: 2026-09-30
- title: "Statement regarding reported activity targeting Government of Canada websites"
url: "https://www.cyber.gc.ca/en/news-events/statement-regarding-reported-activity-targeting-government-canada-websites"
archive_url: "https://web.archive.org/web/20261002213902/https://www.cyber.gc.ca/en/news-events/statement-regarding-reported-activity-targeting-government-canada-websites"
publisher: "Communications Security Establishment Canada"
type: government-advisory # Official statement published on the Canadian Centre for Cyber Security site.
date: 2026-09-29
- title: "OpenAI says its models engaged with US government websites in misbehavior disclosure"
url: "https://www.opb.org/article/2026/09/26/openai-says-its-models-engaged-with-us-government-websites/"
archive_url: "https://web.archive.org/web/20260930224927/https://www.opb.org/article/2026/09/26/openai-says-its-models-engaged-with-us-government-websites/"
publisher: "Associated Press" # AP wire story as published by OPB; carries OpenAI's statement and the Education Department's response.
type: news
date: 2026-09-26
- title: "AI Agents Aimed SQL Injection at US and Canadian Government Sites"
url: "https://www.securityweek.com/ai-agents-aimed-sql-injection-at-us-and-canadian-government-sites/"
archive_url: "https://web.archive.org/web/20261008021256/https://www.securityweek.com/ai-agents-aimed-sql-injection-at-us-and-canadian-government-sites/"
publisher: "SecurityWeek"
type: news
date: 2026-10-02
added:
date: 2026-10-15
by: "MLSecOpsHub"
last_updated: 2026-10-15
139 changes: 139 additions & 0 deletions data/incidents/anthropic-unintended-model-actions-real-websites.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,139 @@
id: anthropic-unintended-model-actions-real-websites
name: "Claude models took unintended actions on real third-party websites during Anthropic evaluations and internal use (2026)"
summary: >-
On 2026-10-09 Anthropic published a review of transcripts, begun in July,
describing cases in which Claude models acted on real third-party websites
and systems in ways it did not intend, mostly during evaluations run with
live internet access and some during internal use. It grouped them into four
categories. First, exploiting basic software flaws: Claude Mythos Preview,
unable to reach a university-hosted scientific tool, used a script on the
university's server to copy files, found an injection flaw in that code and
used it to run commands on the server; Claude Mythos 5 used applications
hosted on a site to accept a data use agreement on its behalf. Second,
submitting forms it should not have: an unreleased research model repeatedly
submitted a real government form when a practice copy failed, and Claude
Haiku 4.5, generating example tasks on randomly selected webpages, submitted
an invented tip about an unsolved homicide through a police department's tip
form. Third, working around restrictions to reach gated data: Claude Mythos 5
reused access tokens from a local government property map's browser settings
file to query its server directly, and used a token issued by a state
agency's public dashboard to query data otherwise offered for a fee. Fourth,
several models, including Claude Opus 5 and Claude Mythos 5, used URL
shortening services to get around fetch-tool URL-length limits. Some cases
involved U.S. federal, state and local government websites; Anthropic
briefed the White House, notified each agency and, at the organizations'
request, named none of them. Anthropic judged the real-world impact minimal
and the cases less severe than its July and September cybersecurity
incidents. The Philadelphia Police Department self-disclosed on the same day
that the tip, dated 2026-07-18, came through PhillyUnsolvedMurders.com, was
flagged as spam and never forwarded for investigation, and involved no
unauthorized access to police systems; it called the two-month delay in
detecting and reporting the incident "unacceptable". Anthropic corrected the
post on 2026-10-12 regarding the purpose of its fetch-tool limits.
date_disclosed: 2026-10-09 # Anthropic report and Philadelphia Police Department statement the same day; Anthropic post corrected 2026-10-12.
status: confirmed # First-party disclosure by the model operator; the one named victim (Philadelphia PD) independently confirmed its case.
confidence: primary
# Graded autonomous-attack rather than lab-escape-eval because real
# third-party systems were acted on, as for anthropic-cyber-evals-real-target-incidents.
category: autonomous-attack
# Anthropic states "minimal real-world impact" and that the cases are
# significantly less severe than its summer cybersecurity incidents. Medium is
# arguable (commands run on a third-party server; a false tip sent to police);
# editor to review.
severity: low
models:
- "Claude Mythos Preview"
- "Claude Mythos 5"
- "Claude Haiku 4.5"
- "Claude Opus 5" # Named in the URL-shortening category only.
- "Anthropic unreleased non-frontier research model (unnamed)" # Government-form case.
model_families:
- claude
actor: "Anthropic evaluation and internal-use agents (Claude Mythos Preview, Claude Mythos 5, Claude Haiku 4.5, Claude Opus 5 and an unreleased research model) acting outside their intended scope"
actor_type: lab-test-eval
autonomy_pct: null
# Most cases occurred in automated evaluation runs; in the internal-use cases a
# researcher set the task but not the workaround the model chose.
autonomy_level: fully-autonomous
guardrail_bypass:
- none-observed # No attacker; Anthropic attributes the behaviour to the models working around restrictions when a task could not be completed as given.
ai_role: load-bearing
related:
- anthropic-cyber-evals-real-target-incidents # Same transcript review; Anthropic contrasts these cases with the July/September incidents.
mitigations:
- "Anthropic extended its removal of live internet access from some high-risk and cybersecurity evaluations to all internal evaluations until its security and monitoring measures are confirmed to catch such behaviour; some public evaluations were dropped, moved offline or rebuilt so their tasks do not reach live websites."
- "Anthropic restricted what models can do with internet-access tools such as web fetch, and built tooling to automatically detect and block these behaviours, which it says now runs on most evaluations and internal agentic use and blocked all the described cases when tested."
- "Anthropic is fixing or removing training environments that reward working around tool restrictions, migrating internal agents to centrally managed infrastructure with containment, minimizing their internet access, and monitoring agent actions with safety classifiers and hierarchical summarization."
- "Per the Philadelphia Police Department, Anthropic terminated the automated testing process responsible for the tip and instituted an additional validation mechanism; the department's human review and vetting of tips limited the impact."
targets:
orgs_affected: null # Anthropic gives no count of affected organizations. TODO: source
records_exfiltrated: null
sectors:
- government # U.S. federal, state and local government websites, per Anthropic.
- education # A university-hosted server.
- emergency-services # Philadelphia Police Department tip form.
countries:
- US # Anthropic: U.S. government agencies; the Philadelphia PD. The university's location is not stated.
geo:
points:
- role: target
basis: victim-location
attributed_by: "Anthropic"
country: US
lat: 37.09
lng: -95.71
label: "United States (federal, state and local government websites, per Anthropic)"
illustrative: true # Country centroid; lab-test-eval records carry target points only.
lifecycle_phases:
- recon
- initial-access # Injection flaw used on a university server.
- execution # Commands run on that server.
- credential-access # Access tokens taken from a site's browser settings file and reused.
- deception-social-eng # Invented homicide tip submitted to police; Anthropic judges the model was producing example content, not trying to mislead.
mappings:
mitre_atlas: []
mitre_attack: []
owasp_llm: []
owasp_asi: []
cve: []
aiid: []
impact: >-
Per Anthropic, minimal real-world impact: commands run on a university server
through an injection flaw; real online forms submitted, including a
government form and a fabricated homicide tip to a police department; fee- or
token-gated public data obtained from local and state government systems;
URL shortening services used to evade fetch limits. To Anthropic's knowledge
no customer data or Anthropic internal systems were involved. Per the
Philadelphia Police Department, the tip was flagged as spam, never forwarded
to its Real-Time Crime Center, and there was no indication of unauthorized
access to police systems or compromise of department data.
sources:
- title: "Investigating unintended model actions in our evaluations and internal use"
url: "https://www.anthropic.com/research/investigating-unintended-model-actions"
archive_url: "https://web.archive.org/web/20261014152739/https://www.anthropic.com/research/investigating-unintended-model-actions"
publisher: "Anthropic"
type: first-party-disclosure
date: 2026-10-09 # Corrected 2026-10-12; the snapshot post-dates the correction.
# Carries the Philadelphia Police Department's statement in full. Police say
# Anthropic notified them on 2026-10-07 and met them on 2026-10-08; Anthropic
# says it shared the finding on 2026-10-08.
- title: "AI model submitted false tip about unsolved murder, Philadelphia police say"
url: "https://6abc.com/post/anthropic-ai-model-submitted-false-tip-unsolved-murder-philadelphia-police-say/19925243/"
publisher: "6abc Action News (WPVI-TV)"
type: news
date: 2026-10-09
- title: "Anthropic AI model submits false homicide tip to Philadelphia police"
url: "https://www.aljazeera.com/news/2026/10/10/anthropic-ai-model-submits-false-homicide-tip-to-philadelphia-police"
publisher: "Al Jazeera"
type: news
date: 2026-10-10
- title: "Claude submitted fake murder tip to Philly police during summertime testing"
url: "https://www.theregister.com/security/2026/10/13/claude-submitted-fake-murder-tip-to-philly-police-during-summertime-testing/5302842"
archive_url: "https://web.archive.org/web/20261014184433/https://www.theregister.com/security/2026/10/13/claude-submitted-fake-murder-tip-to-philly-police-during-summertime-testing/5302842"
publisher: "The Register"
type: news
date: 2026-10-13
added:
date: 2026-10-15
by: "MLSecOpsHub"
last_updated: 2026-10-15
Loading
Loading