.png)
On August 5, 2026, OpenAI disclosed that a collective of AI agents under evaluation had broken out of their sandboxes and taken admin control of the cluster they were running on. It got there, in part, by exploiting Ruby deserialization to execute commands. That caught our attention, because in 2018 we published the first universal RCE deserialization gadget chain for Ruby, built entirely from the standard library with no dependencies. That chain works only against Ruby versions up to 2.6.10, and the most recent public chain only works up to 3.4-rc.
This post releases a new universal chain that turns a single Marshal.load
into command execution on Ruby 4.0.6, the most recent release at the time of writing, and works unchanged as far back as 3.3. The chain is built with new gadgets from untapped sources as well as old gadgets put to new use.
Serialization is the process of converting an object into a series of bytes which can then be transferred over a network or stored on the filesystem or in a database. These bytes include all the information required to reconstruct the original object. This reconstruction process is called deserialization. Each programming language typically has its own native serialization format and may refer to this process by a name other than serialization/deserialization. In the case of Ruby, the terms marshalling and unmarshalling are commonly used, and the operations are provided by Marshal.dump
and Marshal.load
.
Universal Ruby deserialization gadget chains begin in 2018, built on earlier research into application specific chains against Ruby on Rails, and that universal work then fed back into the application specific chains that came after it. Several of the milestones below supply pieces that this chain builds on.
The most recent public chain, published in late 2024, reached command execution on Ruby 3.4-rc with this payload:
Ten days after it was published, two commits landed in RubyGems that removed the gadgets it relied on, each citing the writeup as motivation. Both shipped in Ruby 3.4.0, which is why the chain works against the release candidate but not against the release.
The first commit, 62b49465f8, is titled "Improve type checking in marshal_load methods" and notes that it "Makes it harder to use those classes as gadgets".
Gem::Version#marshal_load
had passed the deserialized value straight to the constructor without validation, where Gem::Version.correct?
calls to_s
on it:
The second commit, 89ad04db86, is titled "Stop storing executable names in ivars" and notes that it "Removes usage of these classes as ACE gadgets".
`Gem::Source::Git`
and `Gem::Resolver::GitSet`
had stored the git executable name in an instance variable, which Marshal restores directly and which was later handed to a process spawn:
The name is now read from the environment at the point of use, so there is no instance variable left to set.
These two commits broke to_s_wrapper
and exec_gadget
, but Gem::SpecFetcher
and call_url_and_create_folder
were left alone and work in Ruby 4.0.
The chain opens with Gem::SpecFetcher
not because the class does any work, but because Marshal.load
has to resolve the constant, and resolving it fires the RubyGems autoload that requires the file defining it, which in turn requires files of its own, and so on. A bare Ruby process therefore starts with a small set of classes reachable by a chain and ends up, after a single constant reference, with a much larger one to pick gadgets from, including Gem::URI::Generic
, `Gem::RequestSet::Lockfile`
and `Gem::StubSpecification`
, all of which the rest of this chain depends on.
A suitable replacement for exec_gadget
is supplied by Gem::Specification.load
, where Gem.open_file
resolves to File.open
:
This method reads a file from disk and passes its contents directly to eval
, so a chain that can control both the filename handed to Gem::Specification.load
and the contents of that file ends up with arbitrary code execution.
The available set offers no flexible gadget of the form @controlled.load(@also_controlled)
, but `Gem::StubSpecification`
provides an indirect route to `Gem::Specification.load(loaded_from)`
by calling the hash
method. This works because loaded_from
is an attr_accessor
, so its value is held in @loaded_from
and can be set through deserialization:
That leaves the question of how hash
gets called during deserialization.
Ruby invokes hash
on an object whenever it is used as a key in a Hash
. Marshal.load
reconstructs a hash by inserting its keys, so placing the crafted Gem::StubSpecification
as a key somewhere in the payload is enough to have hash
called.
Java aficionados will recognise this.HashMap.readObject
callshashCode
on every key it restores, which is the entry point for a large share of the chains in[ysoserial].
The trigger is not a niche marshal_load
override that a maintainer can quietly tighten, but the interaction between two fundamental features of the language, namely hashing an object and reconstructing a Hash
during deserialization. Removing it would mean changing the way core data structures behave, which is exactly the kind of tradeoff where a gadget can be cheap to use and expensive to forbid.
Being able to eval
an arbitrary file on disk is only useful if the chain can also write attacker-controlled code to disk. Rather than build a new primitive for this, the chain reuses call_url_and_create_folder
, which is one of the pieces of the 3.4-rc chain that the maintainers left untouched.
In that earlier chain the gadget created the directories that the command-execution gadget depended on, since Gem::Source::Git
began by changing into one of those directories and would fail if it did not already exist. Here it is put to a different use: its URL-download functionality fetches attacker-hosted content and writes that content onto the filesystem at a predictable and typically writable path by way of directory traversal.
The 3.4 chain invoked call_url_and_create_folder
through to_s_wrapper
, which the type checking commit removed, so the gadget needs a new caller.
It also needs a caller that tolerates failure. The gadget expects the URL it fetches to hold a serialized object and raises when it does not, and what has to land on disk is Ruby source. A polyglot that is valid as both is not possible, because the Marshal header leaves no room for one. The download and the write happen before the parse, so the exception arrives after the useful work is done.
Ruby's own Time
deserialization provides both. time_mload
validates the zone name inside rb_rescue
, which discards any exception it raises:
time_mload
backs Time._load
, which Marshal.load
calls when rebuilding a Time
. The zone name comes straight from the payload, so a crafted Time
puts an arbitrary object into validate_zone_name
. StringValueCStr
then calls to_str
on it rather than to_s
.
Gem::URI::Generic
closes that gap. Its to_str
is an alias of to_s
, and that method calls to_s
on the @port
attribute:
Wrapping the download gadget in one turns the to_str
call into the to_s
call it needs:
Two gadgets died, two survived, and the survivors do a different job in this chain:
The 3.4 chain pointed call_url_and_create_folder
at a real gemspec on rubygems.org. Any valid URL would have done, since only the directory created along the way was wanted. This time the file holds the Ruby code to be executed. The retrieved contents pass through Gem::Util.inflate
before being written to disk, so the file has to be deflated first:
call_url_and_create_folder
sets the @scheme
attribute to s3
to reach the directory traversal in @port
. The signed URL that s3_uri_signer.rb
builds hardcodes https://
, so the file must be served over HTTPS. The destination is controlled by Gem::Source#fetch_spec
, which joins the cache dir with Gem::MARSHAL_SPEC_DIR
(set to quick/Marshal.4.8/
) and the name tuple's spec_name
of "#{name}-#{version}.gemspec"
, which becomes name-.gemspec
because the @name
attribute is set to "name"
and the version is absent. The inflated copy lands at /tmp/quick/Marshal.4.8/name-.gemspec
, which is the path eval_file_gadget
is given.
Every other gadget in the chain is a plain object whose instance variables can be set with allocate
and instance_variable_set
, then handed to Marshal.dump
. Time
is not, because it defines _dump
rather than being dumped field by field. time_dump
writes the real zone of the real Time
object it is given, so there is no way to make Marshal.dump
emit a Time
whose zone is an arbitrary object. While Ruby will not dump such an object, this does not prevent Marshal.load
from accepting one.
One way around the inability to dump a Time
of the required shape is to dump a stand-in object of the same shape and patch the bytes afterwards. The generator builds an Object
carrying two instance variables, @offset_placeholder
and @zone_placeholder
, and then rewrites the object header and the two attribute names into the form Time._load
expects:
The replacement is a TYPE_USERDEF
(u
) entry for Time
holding the eight byte packed time buffer, wrapped in a TYPE_IVAR
(I
) so that the offset
and zone
attributes ride along with it, exactly as a genuine Marshal.dump(Time.now)
would look. The zone
value that follows in the stream is untouched and is still the gadget.
The patch is byte level surgery on a format with backreferences, so it is fragile in one specific way. Marshal writes each symbol once and emits a TYPE_SYMLINK
for every later use, and a symlink is an index into the symbols seen so far. Adding or removing a symbol definition before the patched region would shift every index after it and corrupt the rest of the stream. Both the search and the replacement therefore define exactly three symbols, so the table stays aligned.
One more detail is needed before the stream can be produced at all. The chain places the Gem::StubSpecification
gadget as a Hash
key, and Ruby calls hash
on a key when the hash literal is evaluated, which would fire the gadget inside the generating process rather than the target. Stubbing the method out for the duration of generation avoids that:
Every gadget above plays one of two roles: getting the attacker's code onto disk, or reading it back and running it. Laid out as a single chain, the whole thing looks like this:
Running the generator emits the finished chain:
Two things have to be in place before it is loaded. The deflated payload from earlier must be served as poc-id.rz
over HTTPS by the host named in @host
, which is example.com
here and would be a reachable attacker controlled host in practice. The target must also be able to write to /tmp
, though any writable directory would do if the traversal and the filename are changed together. Nothing else is required of the target: no gems beyond those loaded and available by default in Ruby, no application code, and no prior state on disk.
Running the generated payload against an empty Ruby process using the Docker image ruby:4.0.6
outputs uid=0(root) gid=0(root) groups=0(root)
, showing the id
binary was successfully executed:
The exception that follows is expected and harmless. Gem::Specification.load
has already passed the fetched source to eval
, but that source ends with puts
, so the value it hands back is nil
rather than a gemspec. The method warns and returns nil
, and Gem::StubSpecification#hash
raises when it tries to read a name from it.
It is also avoidable, which matters if a stack trace in the logs or an aborted request is something you would rather not leave behind. Gem::StubSpecification#hash
is name.hash ^ version.hash ^ platform.hash
, and each of those three reads a field off whatever Gem::Specification.load
returned. Ending the evaluated source with a Gem::Specification
is therefore enough for Marshal.load
to return normally, with no warning and no exception:
One side effect is worth noting for anyone reproducing this: the evaluated source is left behind at /tmp/quick/Marshal.4.8/name-.gemspec
.
The chain turns a single Marshal.load
into command execution on Ruby 4.0.6 and works unchanged as far back as 3.3. It needs no gems beyond those that ship with Ruby, no application code, and no prior state on disk. Outside the target process it needs only a reachable HTTPS host and a writable directory.
Little of it had to be built from scratch. The two commits that followed the 3.4 writeup removed the gadgets they named and left call_url_and_create_folder
alone, so it is still here, doing a different job than it did before: fetching attacker-controlled bytes onto disk rather than creating a directory some other gadget depended on. Gadgets outlive the chains they are found in. When a chain stops working, the surviving gadgets can be recycled into the next chain.
What is new is where the rest of the chain comes from. Every public Ruby chain until now has been built entirely out of methods written in Ruby, in the standard library or RubyGems, that a maintainer can tighten in a five-line diff, which is exactly what happened to Gem::Version#marshal_load
. This one reaches below that. The failure-tolerant caller it needs is time_mload
, which is C, and which throws away the exception the download gadget raises because rb_rescue
was there to keep a malformed zone name from breaking Time
deserialization. The trigger is C as well, and is not an override at all but the fact that a Hash
calls hash
on its keys while Marshal.load
rebuilds it. Neither is a stray convenience that can be quietly deleted. Removing them means changing how Time
deserialization tolerates bad input and how core data structures behave, and neither is a change the language can realistically make.
So the advice does not change, but it is worth being precise about why. Removing gadgets raises the cost of writing a chain; it does not remove the capability, because the gadgets are spread across a library that is loaded into every Ruby process by default and are found faster than they are patched. Marshal.load
on untrusted input is command execution, on the current release, with no dependencies. Treat it that way and use a data-only format instead.
This post opened with a collective of AI agents that took admin control of a cluster, in part through Ruby deserialization. Whether they assembled a chain of their own or reused a published one, the assumption that no chain exists for the version in front of you was never a control, and it is no longer even a delay. If you came to this post wondering whether Ruby deserialization is still worth caring about in 2026, the agents that broke out of that cluster have already answered it: yes.
Until we deserialize again, ciao bella!